AI for decisions.

The model that named the category. A hosted API: send your app's state plus typed questions (Choice, Score or Noul) and get a probability back for each answer. It generates no text, and its weights are not public.

Price

$0.042 per million input tokens; output free · Compare prices

Latency

70–500 ms end to end (TypeSafe-reported)

Version

jev-1.13 is the stable release; jev-1.14 has been the preview build since 14 Oct 2026

Context

262,144 tokens on OpenRouter (listing and guide checked 8 Nov 2026); direct-provider limits may differ

Access

Launched in early access; also via OpenRouter

Question types

Choice, Noul and Score; experimental Box questions require Jev 1.14

• 40x–200x faster than frontier LLMs on comparable tasks vendor reportedsource
• Trained with Reinforcement Learning for Calibrated Decisions (RLCD); no technical paper published as of 2 Oct 2026 vendor reportedsource
• Leads JevBench's capability ranking of Jev-class systems (80.0) on 1 Oct 2026 independentsource
• No public weights: TypeSafe's own Hugging Face account listed zero public models when checked on 20 Oct 2026 vendor reportedsource
• The 14 Oct 2026 update switched on experimental image input and opt-in long context for Jev 1.13 and 1.14 (1.13 was "multimodal all along"); bounding-box answers need 1.14 vendor reportedsource
• Tops Artificial Analysis's Decision Index at launch with 58, ahead of GLiDE and Microsoft-Decision-1 on 54 (13 Oct 2026) independentsource

Source: Announcement → Source: Jev 1.14 → Source: Docs → Source: On OpenRouter → Source: Experimental Box primitive →

OpenAI's Decisions API uses GPT-6 Luna to return typed answers for classification, routing and agent actions. It launched in public beta on 6 Oct 2026 and became generally available on 22 Oct. Developers supply the questions and allowed answers; the endpoint returns probabilities, choices or rubric scores.

Model

gpt-6-luna only

Access

Generally available since 22 Oct 2026; public beta launched 6 Oct; native GPT-Live integration since 15 Oct

Price

$0.10 per million uncached input tokens; output free · Compare prices

Inputs

Text, images, JSON and conversation history with tool calls and outputs

Questions

Up to 400 per request, increased from 200 in public beta

Endpoint

POST /v1/decisions

Source: Decisions guide → Source: API changelog →

Liquid AI's decision model family. The hosted d1 takes the same three question types as Jev (Noul, Choice and Score). On 7 Oct Liquid released two small open siblings you can run on a device, and on 15 Oct a bigger hosted one, d1-ultra, joined OpenRouter. On 27 Oct d1.1, the successor to d1, joined OpenRouter at the same price, with -latest aliases that follow each family's newest version. On 27 Oct Liquid announced two more open models, d1-2B (text, pictures and video) and d1-Omni-2B (adds audio), built on LFM3; their Hugging Face repositories went up on 24 Oct.

Hosted

d1.1, d1 and d1-ultra, via Liquid's API and OpenRouter

Open weights

d1-3B and d1-omni-600M (7 Oct), then d1-2B and d1-Omni-2B (announced 27 Oct), under the LFM Open License v1.0

Price

OpenRouter lists d1 and d1.1 at $0.02 and d1-ultra at $0.042 per million input tokens, output free (as of 28 Oct 2026) · Compare prices

• Liquid says d1 is the first model to beat Jev on the Jev Decision Index, a community index hosted on Hugging Face. We could not find a d1 result in the index's public data (release 0.2.1, generated 28 Sep) vendor reportedsource
• d1-3B answers a question in 16 ms on an NVIDIA Jetson AGX Thor vendor reportedsource

Source: Docs → Source: Announcement → Source: Decision Index → Source: Open d1 release → Source: d1.1 on OpenRouter → Source: d1-2B →

Inception's decision model, served as a System One endpoint: send state and typed questions, get a choice, a score or a yes/no answer, each with a probability.

Price

$0.02 per million input tokens at a 50% discount, output free; a free tier remains (OpenRouter, as of 22 Oct 2026) · Compare prices

Context

66K tokens (OpenRouter)

Access

OpenRouter

• Up to 14 decisions per second vendor reportedsource
• The most intelligent decision model on OpenRouter, on JevBench v1.4 (the board was on v1.5.4 as of 2 Oct 2026) vendor reportedsource
• On Vals AI's LegalBench sample, 0.869 to Jev's 0.740 accuracy; on claim verification its router stayed inside a 1% error budget where Jev's went over independentsource

Source: OpenRouter → Source: Announcement →

Upstage's decision endpoint, available through OpenRouter. The date is when OpenRouter announced it; we have not found an announcement from Upstage itself. A faster sibling, Solar Decide Flash, joined on 8 Oct 2026 with a 524K context.

Price

OpenRouter lists $0.05 per million input tokens, output $0 (as of 2 Oct 2026); the launch post described a limited-time 50% discount · Compare prices

Access

OpenRouter

Source: OpenRouter → Source: OpenRouter announcement → Source: Solar Decide Flash →

A one-pass classifier for agent traces. For each behaviour you define, it returns the probability that it is present, absent or not observable, without generating text. Respan's later Span-02 (15 Oct 2026) reads images too, can return a number directly, and ships as Apache-2.0 open weights.

Price

$0.02 per million input tokens; output free · Compare prices

Free tier

Span-01 Lite

Maker

Respan, a Keywords AI product

• Beats GPT-6 Luna and Jev 1.13 on Respan's own behaviour-detection benchmark vendor reportedsource

Source: Announcement → Source: Span-02 → Source: Span-02 weights (Apache-2.0) →

A hosted decision model served only at /v1/systemone, the request format TypeSafe's SDK speaks. Its site doesn't say who runs it or when it launched.

Price

$0.03 per million input tokens; output $0 · Compare prices

Endpoint

/v1/systemone

• 0.768 accuracy against Jev's 0.727 on typed-decision benchmarks vendor reportedsource

Source: meragpt.com →

Microsoft's decision-scoring model for routing, classification, prioritisation, verification and workflow control. Post-trained from Qwen3.5-9B, available in Microsoft Foundry and on OpenRouter.

Price

$0.042 per million input tokens; output free · Compare prices

Base

Qwen3.5-9B

Access

Microsoft Foundry, OpenRouter

• Ties GLiDE for second on Artificial Analysis's Decision Index at launch, on 54 (13 Oct 2026) independentsource

Source: Announcement →

The Clef family on Workers AI answers typed questions in one pass. Clef-omni adds audio and video, so it can decide about what is seen and heard at the same time. The same update cut Clef-flash to $0.038 per million input tokens.

Size

30B mixture of experts, 3B active

Price

$0.150 per million input tokens · Compare prices

Access

Cloudflare Workers AI; Clef-flash weights are on Hugging Face

Source: Cloudflare changelog →

A decision model you call from SQL. It classifies, scores, filters and routes rows inside Snowflake through AI_COMPLETE. A lower-latency Decision Fast, with REST access, followed on 19 Oct.

Access

Private preview in Cortex AI Functions

Limit

Up to 32 options per question

• Scored 57.63 on a subset of the Hugging Face decision benchmark (the 32-option limit rules out the full suite) vendor reportedsource

Source: Announcement → Source: Decision Fast →

A closed decision model for text, structured data and images that can give a one-sentence reason alongside each probability.

Price

$0.04 per million input tokens; output free · Compare prices

• 80.0% on jev-bench at a 67 ms median response, ahead of Jev 1.13 and Mercury Decide on its own run vendor reportedsource

Source: Announcement →

Subquadratic reports a 384K-token context for text and images. Snap 1.2's reported price is $0.04 per million input tokens. Subquadratic's 11 Nov 2026 release note says requests using retired subq-snap-1.1 and subq-snap-preview IDs route to 1.2, and advises retesting thresholds because decisions and probabilities may change.

Price

$0.04 per million input tokens; output free · Compare prices

Inputs

Text and images; vendor-reported 384K-token context

Access

Subquadratic's API; the Snap 1.1 Baseten listing is a separate route

API model

subq-snap-1.2; subq-snap-latest alias. Retired subq-snap-1.1 and subq-snap-preview IDs route to 1.2

• Subquadratic's 22 Oct 2026 release post quotes an Artificial Analysis Decision Index score of 65 for SubQ Snap 1.1, tied for #1. This figure is quoted by the vendor. vendor reportedsource
• Subquadratic says Snap 1.2 choice results are identical across option orderings. The release note gives no numeric accuracy improvement. vendor reportedsource

Source: Announcement → Source: Version 1.1 release → Source: Baseten → Source: API models, status and pricing → Source: Baseten launch, 02 Nov 2026 → Source: Snap 1.2 release and retired IDs →

Scores text or a whole document against your own choice, yes/no and ordered-scale questions. Request-compatible with Jev's Decisions API. The date is OpenRouter's listing.

Context

1M tokens (OpenRouter)

Access

OpenRouter, ScaleDown API

Source: Docs → Source: OpenRouter →

A new branch of the family: a System One extraction model. Where a decision model picks from answers you list, Clairic extracts, computes, estimates or forecasts values from unstructured input and returns them as typed answers: an invoice total, a delivery date, a repair cost, a 24-hour traffic forecast. It does this quickly, without an LLM writing prose.

Access

Workers AI as @cf/cloudflare/clairic; back on OpenRouter as cloudflare/clairic when checked 4 Nov 2026

Weights

Cloudflare says open weights are coming soon (as of 19 Oct 2026)

Price

$0.10 per million input tokens on OpenRouter; output free · Compare prices

Inputs

Text and images

Context

65,536 tokens on OpenRouter

• Higher accuracy, lower latency and lower cost than traditional LLMs at extraction vendor reportedsource

Source: Announcement → Source: OpenRouter listing → Source: OpenRouter decisions inventory (34 entries, checked 4 Nov 2026) →

A decision engine that searches the live web before it answers. You send typed questions about a company, product or policy, and one call returns typed answers with probabilities (or a refusal) plus a shared list of the web pages it used, so it can answer things a model only knows from training data cannot, such as whether a company has an open job right now.

Access

Hosted API, with a playground; 5,000 free requests to start; Decisions Fast also on OpenRouter from 6 Nov 2026, with a seven-day free promotion

Price

$1 per 1,000 calls (as of 25 Oct 2026) · Compare prices

Grounding

Searches the live web per request; citations cover the whole call, not each answer

Latency

VENDOR-REPORTED: 0.56 seconds median on PWS-Decisions, measured by Parallel on 29 Oct 2026 (Pacific).

• Parallel says it costs 10x less than a standalone model vendor reportedsource
• Parallel reports 76.22% on its own 300-question PWS-Decisions set with live web search, against Perplexity Decider v1.2 at 50.89%, OpenAI's Decisions API at 41.67% and Jev 1.14 at 33.44%, run without web access or supplied external context. Uniform random guessing is 42.42%. LLMs with native search scored 96.56% to 97.33% at about 16.5 to 28.8 times the cost. Parallel generated these questions from its own customer topics; dataset terms prohibit redistribution. vendor reportedsource
• The benchmark blog chart labels Baseten Elephant at 32.4% overall and 32.3% public. The repository records Baseten Mammoth at 45.0% overall and 37.0% public. vendor reportedsource

Source: Announcement → Source: Docs → Source: Benchmark → Source: Benchmark announcement → Source: OpenRouter launch and promotion → Source: OpenRouter decision inventory →

Kev is Jared Palmer's family of open, Jev-compatible decision models, from 0.8B to 27B parameters. A pointer head on a Qwen base scores each option, and the API matches TypeSafe's, so their SDK works against a local Kev server. Kev 1.0, released 1 Oct 2026, pins the family's checkpoints, model cards, evaluation suites and serving code; it introduces no newly trained weights.

Sizes

0.8B, 4B, 9B, 27B

Licence

Apache-2.0

Base

Qwen3.5 (27B: Qwen3.8)

Release

Kev 1.0, 1 Oct 2026

Stars

9,758 on GitHub (GitHub API, 2 Nov 2026)

Price

$0.042 per million input tokens and $0 per million output tokens for jaredpalmer/kev-4b on OpenRouter (VENDOR-REPORTED, 2 Nov 2026) · Compare prices

Validated context

VENDOR-REPORTED: 65,536 tokens for Kev-27B; 8,192 for 0.8B, 4B and 9B, measured on CUAD contracts

• Kev 1.0's breadth-v1 held-out-dataset test scored a chance-corrected index of 23.3 for 0.8B, 38.0 for 4B, 41.0 for 9B and 52.3 for 27B, against Jev's 54.0, in the maker's fp32 evaluation at shipped temperatures. No Kev trained on these datasets. vendor reportedsource
• Kev-27B's 65,536-token validated context is the longest tested state whose CUAD accuracy stays within three percentage points of its 8k-token accuracy at the 95% lower bound. The smaller models already miss that tolerance at 16k. vendor reportedsource

Source: GitHub → Source: Kev 1.0 release → Source: Hugging Face → Source: Kev 4B on OpenRouter → Source: GitHub API: stars →

A local decision model built on ModernBERT-large: give it a state, a question and options, and it returns a probability for each option in one encoder pass. It handles choice and yes/no questions without generating text, through PyTorch, MLX or ONNX. The open weights are under Apache-2.0. The date is when the weights repository went up on Hugging Face.

Size

395M parameters, ModernBERT-large plus a 4 KB scoring head

Licence

Apache-2.0

Context

8,192 tokens; longer sequences are refused

• The maker reports 29.57 on Decision Index 0.2.1 (raw 46.80) from one complete run, with a second seed scoring 29.13 (raw 45.96). These are the maker's runs, not an independent reproduction. vendor reportedsource

Source: Hugging Face →

Together's Jev-inspired experiment: a Qwen3.5-4B fine-tune that reads a state, a question and 2 to 24 lettered options. Together says plainly it is not a non-autoregressive runtime like Jev: it keeps a standard next-token head and generates its answer as one letter.

Base

Qwen3.5-4B

Licence

Code MIT; weights licence being finalised as of 2 Oct 2026

Run it

Together serverless, or Ollama

• Together reports 880/1,000 (88.0%) and 300/300 (100%) on evaluation sets used during development (VENDOR-REPORTED). vendor reportedsource
• In jev-rsi's independent baseline test, the author measured untuned Qwen3.5-4B using Tev1's decision protocol through local Ollama at 673/1,000 (67.3%) on test and 130/300 (43.3%) on policy_transfer. independentsource

Source: Hugging Face → Source: GitHub → Source: Announcement → Source: Untuned 4B baseline test →

A 340M-parameter encoder decision model from the team behind GLiNER, fine-tuned from GLiNER2. It scores the allowed answers to typed questions in one pass and is small enough to run on a CPU.

Size

340M parameters

Licence

Apache 2.0

Latency

167 ms p50 on a Xeon CPU, 38 ms on a V100 (Fastino-reported; batch 1, 64-token input)

• Best average across 17 datasets on Fastino's own benchmark: 60.1% in the launch post, 60.2% on the model card as of 2 Oct 2026 vendor reportedsource

Source: Announcement → Source: Hugging Face →

A non-autoregressive decision engine on ModernBERT-large (421M), with a 322M multilingual sibling. Its author, Nandakishor Mukkunnoth, says he built the same approach a year before Jev launched.

Size

421M (English), 322M (multilingual)

Licence

Apache 2.0

Latency

32.8 ms on a single GPU (maker-reported)

• 6 to 8 times faster than Jev vendor reportedsource
• Built and open-sourced a non-autoregressive decision model a year before Jev vendor reportedsource

Source: Site → Source: GitHub → Source: Model card and preprint citation →

A 9B decision model, LoRA-tuned from Qwen3.5-9B, with its 2,676 training examples published. One of the first two models in Ollama's decision support. Bespoke Labs followed with hosted Nimble v4 (9B) and v4-max (27B) on 13 Oct 2026.

Size

9B

Licence

Apache 2.0

Run it

ollama pull nimble

• Matches 90.1% of reference labels on 324 held-out examples, against 93.2% for Jev 1.13 vendor reportedsource

Source: GitHub → Source: Ollama → Source: Hugging Face → Source: Nimble README →

Open decision models from the vLLM Semantic Router project, published as "Decision 1.0: towards open decision foundation models". Six were announced on 22 Sep. The follow-up, Decision 3.0 (d3), is Apache-2.0 and reads text, images and video, with sizes up to 27B.

Sizes

0.6B to 9B

Licence

Apache-2.0 (checked on Nox-4B)

Source: Hugging Face → Source: Announcement PR → Source: d3 on Hugging Face →

Perplexity's decision model reads text and images and returns a probability for every answer. It ships as Apache-2.0 open weights and as a hosted API. V1.1 (27B) halved the price on 6 Oct; V1.2 and a 4B Decider Flash followed on 20 Oct.

Latest

Decider V1.2 (27B) and Decider Flash V1.2 (4B)

Price

$0.02 (V1.2) and $0.01 (Flash) per million input tokens on OpenRouter, output free (as of 22 Oct 2026) · Compare prices

Licence

Apache-2.0

• Decider V1.1-27B was the most accurate of seven decision models (94.7%) on 799 unpublished yes/no questions about coding-agent work and destructive shell commands, and the only one to pass SILT's registered calibration rule. Jev's result was inconclusive. So were the results for OpenAI's Decisions API and Laya (11 Oct 2026) independentsource

Source: V1.2 announcement → Source: Hugging Face → Source: V1.1 announcement →

Drex 2.0 returns probabilities over typed answers for text, with open weights under Nace.AI's Open RAIL licence. Drex Max (13 Oct) adds images and screenshots; its 27B weights are now public under Nace.AI Open RAIL-M. Drex Max Finance followed on 01 Nov 2026 as a hosted model. The entry's date is Drex 2.0's release in Nace.AI's model table.

Size

Drex 2.0: 35B, about 3B active; Drex Max: dense 27B

Price

$0.03 per million input tokens for Drex 2.0, Drex Max, Max v1.1 and Max Finance, hosted at Nace · Compare prices

Context

Drex 2.0 state up to 131,072 tokens; Max Finance whole request up to 262,144 tokens

Licence

Drex Max weights: Nace.AI Open RAIL-M

Access

Drex Max supports text and images through Python on NVIDIA GPUs. CPU and Apple Silicon runtimes support text only. Max Finance API id: drex-max-finance.

• Nace.AI reports 85.7% accuracy across 3,361 decisions on 48 held-out finance benchmarks for Drex Max Finance, with a Finance Index of 77.3. These are its own 02 Nov 2026 launch figures. vendor reportedsource

Source: Model table → Source: Hugging Face → Source: Drex Max weights and local inference → Source: Drex Max Finance API guide → Source: OpenRouter decision inventory →

A 4B open model built on Qwen3.5-4B that runs on unmodified vLLM. Version 1.2.2 takes up to four images per request.

Size

4B

Licence

Apache-2.0

• Its text-only v1.1 ranked #1 of 104 open-weight systems on JevBench v1.6.1's composite, 72.5 to Jev 1.13's 71.5, read 7 Oct 2026 vendor reportedsource

Source: Announcement → Source: Hugging Face →

Open decision models that read option-label probabilities in a single pass with thinking switched off: you get a pick and a confidence, not an essay. GGUF and MLX builds sit next to the full weights. Released 10 Oct 2026, written up on 16 Oct.

Sizes

9B and 27B

Licence

Apache-2.0

Source: Release post →

A 27B open decision model that answers in English, Hindi, Tamil and Kannada. The makers say Jev still leads on some reasoning benchmarks.

Size

27B, full fine-tune

Licence

Apache-2.0

• 90.7% vs Jev's 83.9% on the maker's held-out decision set vendor reportedsource

Source: Release post →

IBM's switchable models do chat, decisions and speech-to-text from one endpoint. 4.1 adds a general-purpose Decision adapter that speaks the /v1/systemone interface.

Sizes

8B and 30B

Licence

Apache-2.0

Source: Announcement →

A decision model for speech. It listens to the microphone and your text-to-speech at the same time and keeps deciding whether to speak, keep listening, yield or drop in an "mm-hm", scoring audio in 80 ms frames streamed in 320 ms hops. Built for full-duplex voice agents, with an open-source agent (tyger-s2s) on top.

Base

NVIDIA Nemotron 3.5 streaming ASR 0.6B

Licence

OpenMDW-1.1

Works with

Any LLM and TTS (maker's claim)

Source: Hugging Face → Source: Announcement → Source: Voice agent →

A decision model that stands between AI agents and the controllers that run factories. Every Modbus or OPC UA request goes through one gate, where a small model decides in a single pass to allow it, answer from cache, defer it or deny it, and every decision is logged. The date is when the repo and weights went up.

Size

0.8B, built on Qwen3.5-0.8B

Licence

Apache-2.0

• About 5 ms per decision on a laptop CPU vendor reportedsource

Source: GitHub → Source: Hugging Face →

Two open decision models from Base Labs, Baseten's research team: mammoth-1 (27B) and elephant-1 (9B), post-trained from Qwen3.5. They read text and images and return a probability for every option of each question, with all questions sharing one prefill of the state. Baseten launched them claiming the top spot on the community Jev Decision Index. On 29 Oct a third, smaller model joined: joey-1 (2B, built on Qwen3.5-2B), pitched as the fastest and cheapest.

Sizes

mammoth-1 27B (Qwen3.5-27B), elephant-1 9B (Qwen3.5-9B) and joey-1 2B (Qwen3.5-2B), 16,384-token context

Licence

MIT

Price

OpenRouter lists mammoth-1 at $0.10, elephant-1 at $0.035 and joey-1 at $0.01 per million input tokens, output free (as of 30 Oct 2026) · Compare prices

• Baseten says mammoth-1 ranked #1 of 208 models on Jev Decision Index 0.4 with a full score of 71.24, and elephant-1 is the best decision model under 15B parameters (62.65) vendor reportedsource
• In the index's current public data (Decision Index 0.4.1, generated 22 Oct 2026), mammoth-1 is first of 228 on both public-benchmark scores (75.34 skill, 81.19 raw) but third on the full score (68.44), behind GLiDE v1.2 (71.52) and jeffan-31b (69.75) independentsource
• Baseten says joey-1 has a P50 latency of 20 ms, 5x faster than comparable decision models vendor reportedsource

Source: Announcement → Source: Technical report → Source: mammoth-1 on Hugging Face → Source: elephant-1 on Hugging Face → Source: joey-1 →

A tiny open decision model that runs locally, built on Qwen3.5-0.8B and reading text and images. The date is when the weights went up on Hugging Face.

Size

0.8B, built on Qwen3.5-0.8B-Base, about 1.3 GB of RAM on a Mac in 8-bit (maker's figure)

Licence

Apache-2.0

Runs on

llama-server (GGUF), MLX on Apple silicon, or a System One API

• First of 27 decision models of around 1B parameters on the public split of Benchmark Heaven (77.1%), and the fastest at a median 3.5 ms per request (batch size 1, one NVIDIA L40S), in the maker's own comparison vendor reportedsource

Source: Model page → Source: Hugging Face →

Six English document-decision packages share the decido API. Decido-1, Max and Think use Gemma-4-26B-A4B; Fast uses Qwen3-1.7B, Ultra Qwen3-0.6B and XL DeepSeek-V4-Flash. Max distils a three-model research mixture into one checkpoint and generates up to 1,024 reasoning tokens before reading option probabilities. The Max package’s GPU serving path is still being validated.

Variants

Decido-1, Fast 1.7B, Ultra 0.6B, Think and Max 26B-A4B, XL

Inputs

English text or JSON; choice, noul and score

Licence

Apache-2.0 for the five Gemma/Qwen packages; MIT for the XL adapter, with separate base-model terms

Runtime

decido-ai package; GGUF and bf16 packages vary by variant

Speed

Maker reports M4 Pro CPU medians of 4.1 s for Decido-1, 0.64 s Fast and 0.41 s Ultra for a roughly 1k-token state and five questions. Think takes 19 s per question. Max reports 10.1 s p50 for five parallel questions on an A100 80 GB, or about 11 s per question on M4 Pro Metal and 28 s on CPU; no full CPU/Metal accuracy run is supplied.

• The maker reports six-group means of 85.38 for Decido-1 on CPU, 73.01 Fast and 62.16 Ultra. The OpenJev group uses CUDA with the same GGUF, checked against a smaller CPU sample. Training includes public train splits from scored datasets; Jev's published reference run uses a different harness and 19 OpenJev tasks against Decido's 18. vendor reportedsource
• Max reports a six-group mean of 89.23 (95% interval 88.25 to 90.15), against a published Jev 1.13 reference at 83.85. Training uses train splits from scored datasets; Jev is zero-shot and was not rerun in the same harness. Decido uses 18 OpenJev tasks against the reference’s 19. The in-domain comparison is not a general model ranking. vendor reportedsource
• Repeating 50 GPU requests changes Max’s top answer 8% of the time. The card reports identical repeats on CPU and Metal, with answer agreement against GPU on 15 CPU and 101 Metal questions. Those small checks are not complete local accuracy evaluations. vendor reportedsource
• Think’s reported 88.71 six-group mean is rescored from its outputs inside the Pro mixture. For 157 Kev questions the router never sent to Think, Decido-1 answers are used. The reasoning cap is hit on 46% of OpenJev questions. A 60-question CPU sample agrees with recorded GPU choices 98.3% of the time; no full CPU accuracy run is supplied. vendor reportedsource

Source: Decido-1 card and weights → Source: Fast card and weights → Source: Ultra card and weights → Source: Think card and weights → Source: Max card and weights → Source: XL card and weights →

The small decision model inside DocsGPT, used for quick judgements such as whether a message is a prompt injection or whether a chunk answers the question. It runs on CPU with ONNX Runtime. The date is when v0.2 went up on Hugging Face.

Sizes

arc-decide-0.4b-v0.2 for CPU; arc-decide-0.6b-v0.1 for NVIDIA GPUs with vLLM

Licence

MIT, built on Qwen3 (Apache-2.0)

Source: Docs → Source: Hugging Face →

Open Qwen3.5-based decision models for text, images and video. The 1.5 family adds 9B and 27B sizes alongside 0.8B, 2B and 4B. Both Jev System One and OpenAI Decisions request formats are supported. Native 1.5 weights were uploaded on 9 November; the original family dates to October.

Sizes

1.5: 0.8B, 2B, 4B, 9B and 27B; native, ONNX and GGUF formats

Licence

Apache-2.0; licence file included

Inputs

Text, images and video, with backend-specific limits

Runtime

Custom Valen plugins for native GPU serving; ONNX SDKs; custom GGUF runtime. Official unmodified llama.cpp/Ollama support is not claimed. Prebuilt GGUF 0.1.0 lacks direct video.

Training data

The 1.5 training data is not public; the earlier general-training dataset does not establish 1.5 data availability

Demo

The free API serves 1.5 in 0.8B, 2B and 4B sizes, according to the current card

• The maker reports public Decision Index text/vision scores of 69.01/77.85 for 1.5-4B and 74.66/82.91 for 1.5-27B. These are normalized public-track scores, not accuracy percentages or sealed-board results. The published result package records per-size runs. vendor reportedsource

Source: Site and demo → Source: Hugging Face → Source: 1.5 card and weights → Source: 1.5 public result records →

An open decision model for images: give it a photo, document, receipt, chart or screenshot plus a typed question and it returns a probability for every allowed answer, served over Jev's System One API. The date is when the weights went up.

Size

4B, built on Qwen3.5-4B

Licence

Apache-2.0

• Scores 80.6 on the maker's own 2,360 held-out multiple-choice questions from public benchmarks, scored with ImageJevBench's formula, ahead of H2O-Lightning-4B (79.9), Imajev-4B (79.4) and Jev-Vision 8B (75.7); the maker notes this is not the ImageJevBench leaderboard vendor reportedsource

Source: GitHub → Source: Hugging Face →

An open decision model that listens instead of reading: give it an audio clip and a typed question (how does the speaker feel? does the pitch rise?) and it returns a probability for every option, without a transcript. It is Qwen3-Omni-30B-A3B with frozen weights and a decision head in place of the text output. The date is when the weights went up on Hugging Face.

Base

Qwen3-Omni-30B-A3B-Instruct, frozen, with a decision head and 17M parameters of LoRA

Licence

Apache-2.0

Inputs

Audio clip plus Noul, Choice or Score questions

• 73.9 on VocalGrad against 63.0 for Gemini 3.8 Flash, and 83.1 on MMSU against Gemini's 85.7. Cost per 10,000 MMSU questions: an estimated $0.28 of batched GPU time on one H200 at $4.54/hour, against $14.80 in Gemini API charges at promotional pricing vendor reportedsource

Source: GitHub → Source: Hugging Face → Source: Show HN →

A decision model for live audio. You ask yes/no questions in plain English (is music playing? is someone speaking?) and it answers every 40 ms of a stream with yes, no or can't tell, real time on a laptop CPU. The date is when the weights went up on Hugging Face.

Size

About 97M parameters on the audio, plus a 0.6B text encoder (Qwen3-Embedding-0.6B) that reads each question once

Licence

Apache-2.0, code and weights

Inputs

A live audio stream plus yes/no questions; answers yes, no or cant_tell per 40 ms

• Says yes a median 0.24 s after a sound starts over the sounds it caught on 500 real AudioSet clips (speech 0.21 s, 97% of 181 speech onsets caught); reworded questions it never trained on scored 0.892 mean AUROC against 0.911 on its own wording vendor reportedsource

Source: GitHub → Source: Hugging Face →

Fastino’s decision-model family spans 0.4B, 0.8B, 2B, 9B and 27B releases. Each returns option probabilities without generated reasoning. Inputs and serving behaviour vary: the 0.4B encoder and 9B release accept text or JSON and reject images; the original 27B entry supports images. Public weights use Fastino’s community licence, and Fastino also hosts a System One API.

Base and size

27B checkpoint: Qwen3.8-27B, 52.2 GB of bf16 weights, needs one 80 GB GPU. 0.4B: an Ettin encoder derivative, 0.81 GB of weights (1.77 GB tested peak GPU memory). 9B: 15.90 GB of weights (18.92 GB tested peak)

Licence

Fastino community licence (not a standard open-source licence)

Access

Weights on Hugging Face, or Fastino's hosted System One API

Family

Five sizes announced 5 Nov 2026: 0.4B, 0.8B, 2B, 9B and 27B

Inputs

Checkpoint-dependent: 0.4B and 9B accept text or JSON and reject images; original 27B supports images

• Fastino says GLiDE 1.2 is #1 of 229 models on Decision Index 0.4.1 (as of 22 Oct 2026) vendor reportedsource
• In the index's public data (release 0.4.1, generated 22 Oct 2026), GLiDE 1.2 is first of 228 entries on the full score (71.52), ahead of jeffan-31b (69.75) and mammoth-1 (68.44); mammoth-1 leads both public-benchmark scores. In the index's 29 Oct data, GLiDE is second of 247, behind KnowLine 31B Gen1 (71.86) independentsource

Source: Hugging Face → Source: API docs → Source: Decision Index → Source: Family announcement → Source: 0.4B card and weights → Source: 9B card and weights →

A 380M-parameter decision model for data curation: give it a web page, prompt, maths problem or library record and typed questions (topic, format, educational value, difficulty) and it returns a probability for every option. The release centres on its recipe: the data-building and training scripts are public, with source and labeller/licence metadata where known, and 516,964 of the 535,730 training rows are released (web text omitted).

Size

380M parameters, UltraBERT-380M encoder plus a GLiClass scoring head

Licence

Apache-2.0

Open data

516,964 of 535,730 training rows released, plus 1,020,000 FineWeb pages scored with it

Source: Hugging Face → Source: Recipe →

A fast-moving family of open decision models from PelaAI, trained in an agent-driven data loop, with English and Chinese in mind: from 0.8B up to 35B-A3B, with new generations posted every few days. The newest big one, KnowLine-31B-Gen1 on a Gemma 4 base, went up on 23 Oct. The date is the first family release on Hugging Face (4B-Gen1).

Sizes

4B through Gen5.5; 26B-A4B through Gen2.5; 35B-A3B through Gen1.5; 0.8B, 2B, 9B and 31B (Gemma 4 base) at Gen1; community GGUF builds

Licence

Apache-2.0

Languages

English and Chinese

• In the Decision Index's 29 Oct 2026 data, KnowLine 31B Gen1 is first of 247 on the full score (71.86), ahead of GLiDE 1.2 (71.52) independentsource
• PelaAI reports 0.4% of answers changed by an instruction planted in the state, 5.3% with injection templates unseen in training vendor reportedsource

Source: Hugging Face → Source: Decision Index →

Ollama added native decision models through /v1/systemone, based on TypeSafe's Jev API, so you can run one on your own machine. It shipped with Nimble and Tev1.

Endpoint

/v1/systemone

Question types

choice, noul, score

Launch models

Nimble, Tev1

Source: Release notes →

Reproduces Jev's interface pattern with open models: it reads typed option probabilities directly from a model, with no answer sentence and no decoding loop. Runs in your browser, or at home on a GPU, Apple Silicon or CPU. Independent of TypeSafe, and it does not reproduce Jev's model or training.

Availability

Runs in the browser

Models

Qwen3 0.6B, MiniCPM5 2B or Qwen3.5 4B, as pinned GGUF builds through wllama

• Checked against its own README on 2 Oct 2026 code verifiedsource
• On JevBench v1.8.6, published 01 Nov 2026, SemIf (formerly OpenJev, Qwen3.5-4B) ranks #143 at 55.4 Capability, with 27.1 Intelligence and 83.6 Calibration. Jev 1.13.0 is the unranked reference at 77.1 Capability. independentsource

Source: openjev.com → Source: TheoLeeCJ/SemIf-OpenJev →

Parallel constrained decisions for local models through mlx-lm on Apple Silicon: the schema is a set of typed fields, the context is prefilled once and broadcast across one batch row per field. MIT licensed. Large answer sets may be chunked rather than done in a single pass.

Availability

Open source

• Checked against its own code and README on 18 Sep 2026 code verifiedsource

Source: GitHub →

A browser and local playground for decision experiments on memory, uncertainty and control, using Qwen-family models and returning option probabilities. MIT licensed. Its README states it is independent of Jev, of openjev.com and of zhihz/openjev.

Availability

Open source

• Checked against its own code and README on 18 Sep 2026 code verifiedsource

Source: GitHub →

Gemma 3 270M with a scoring head, producing a decision in a single forward pass with no text generation. Code and adapter weights are publicly readable, but the repository carries no licence, so it is not open source in the legal sense. The Gemma base weights remain under their own separate terms.

Availability

Source available; no licence found as of 18 Sep 2026

Model

270M (Gemma 3 backbone)

• Checked against its own code and README on 18 Sep 2026 code verifiedsource

Source: GitHub →

Groot 1 S, a 149M-parameter open typed-decision model that speaks the System One request format and reads states of up to 64k tokens in the maker's tests. Its forward pass also ships in the Bend language.

Size

149.3M

Licence

Apache-2.0

• 75.2% vs Jev's 73.7% on the full 2,000-question test split of the public LocalLLaMA/typed-decisions set, run by the maker, with a Brier score under half of Jev's vendor reportedsource

Source: GitHub →

A self-hosted decision model with a Jev-compatible API: you run the server on your own GPU, and a 230 MB vision adapter on Qwen3.5-4B reads text and images. Python SDK, server and fine-tuning in one package.

Base

Qwen3.5-4B

Licence

Apache-2.0 (code)

Source: GitHub → Source: Hugging Face →

A LiteLLM-based router with an optional Jev decision backend: when a TypeSafe API key is configured it can use Jev to pick a model per request, and otherwise falls back to its own rules-based decider. MIT licensed.

Availability

Open source

• Checked against its own code and README on 18 Sep 2026 code verifiedsource

Source: GitHub →

Runs the open Laya decision models natively on Apple Silicon with MLX: no PyTorch, no cloud API, zero output tokens. Its demo plays Snake, calling Laya for every move with a safety layer that can override unsafe ones. Apache-2.0, and more starred than any project on our list except browser-use's jev-ultrafast.

Stars

6,994 on GitHub (GitHub API, 22 Oct 2026)

Licence

Apache-2.0

• 13.4 ms median for one short English typed decision, 7.4 ms with the multilingual checkpoint, on the maintainer's M3 Max vendor reportedsource

Source: GitHub →

Hume's hosted model answers questions about speech and sound with probabilities. Send one MP3 or WAV recording, 5 seconds to 5 minutes long, with questions in plain language. It supports yes/no, choice, score and time-span answers, and returns no transcript. Hume says the probabilities estimate how human raters would respond.

Access

Playground; request API access through Hume's waitlist

Inputs

One MP3 or WAV per request, 5 seconds to 5 minutes; the model hears 16 kHz mono audio

Price

$0.47 per million input tokens (audio, questions and context); answers not billed. Hume's billing docs, checked 03 Nov 2026 · Compare prices

SDKs

Python hume-audio-decisions and TypeScript @humeai/audio-decisions; Python SDK code MIT

Questions

Several questions can share one recording; the overview demonstrates yes/no and score questions in one request

• Hume reports 70.7% agreement with a human panel's average judgment, against 56.5% for an individual-rater baseline, in its initial evaluation. vendor reportedsource
• On Hume's product-page chart, its model averages 0.82 across 10 benchmarks at 0.23 seconds median latency per screen. The chart gives Gemini 3.8 Flash 78%, Qwen3.8-Omni-Flash 76% and Inkling-Small 67%; it names MMAU, ParaPair and EmphAssess among the benchmarks. vendor reportedsource

Source: Announcement → Source: Product and playground → Source: Docs → Source: Billing → Source: Audio preparation → Source: Launch changelog → Source: Python SDK →

Zoom's hosted Decisions API and Decisions API Pro accept text, images or both. Define several questions about the same input and get structured answers in one call for condition checks, option selection and scoring.

Models

Decisions API and Decisions API Pro

Price

$0.042 per million tokens for Decisions API; $0.10 for Pro, per Zoom's product page checked 03 Nov 2026 · Compare prices

Endpoint

POST /v2/aiservices/decisions/decide

Access

Zoom AI Services portal and playground; sign-up or login required

• Zoom reports 83.9% overall accuracy for Decisions API Pro on a comparison it says uses JevBench and ImajevBench. Its product page says Pro leads the tested models but gives no dated per-model results table. vendor reportedsource

Source: Product, playground and pricing → Source: Launch changelog →

Ant Digital Technologies' System One model for choosing answers to business questions, including native multi-select and decisions ordered by dependencies. IT之家 reported its 22 Oct 2026 launch in 4B and 27B versions for customer service, risk and finance. Ant's current DTMaaS listing describes the lingdt-decision-1.0 endpoint as a 4B test model.

Sizes

4B and 27B announced, per IT之家; the current DTMaaS endpoint identifies itself as 4B

Try it

Open the DTMaaS model page and its API guide. Use an API key with POST maas-api.antdigital.com/v1/systemone and model lingdt-decision-1.0; send state, questions and optional DAG dependencies

Limits

Ant lists 32 questions per request, 255 options, DAG depth 3 and a default 60 requests per minute. Its English and Chinese descriptions disagree on whether the 16K context limit applies per question or per request.

Price

Ant lists ¥0 per million input and output tokens, with an original input rate of ¥0.20 per million tokens. Checked 04 Nov 2026. The rate is for the 4B test endpoint during a free trial; no trial end date was established. · Compare prices

Compare

Jev and Microsoft-Decision-1 are hosted alternatives, each listed at $0.042 per million input tokens. LingDT's free-trial price covers a 4B test endpoint; these rates do not compare accuracy

• Across 23 open benchmarks, Ant says LingDT-Decision-1.0 beats Microsoft-Decision-1 and Jev 1.13, with the 27B version most accurate in its size class. These are Ant's results as reported by IT之家, not an independent comparison. vendor reportedsource
• IT之家 reports Ant's claim of over 90% fewer tokens and up to 10x faster responses than LLMs in its customer-service, risk and finance tests. vendor reportedsource

Source: DTMaaS model page → Source: Official model data and API example → Source: API guide → Source: Launch report (IT之家) → Source: Jev pricing source → Source: Microsoft-Decision-1 source →

A Russian-first typed-decision model built on mmBERT-base (307M), with a separate scoring head. It runs on CPU and supports yes/no, choice and ordinal scales. The maker says it is unaffiliated with TypeSafe AI and was not trained on Jev outputs. The v1.3 card documents improved option-wording robustness and withdraws an earlier calibration comparison.

Size

307M (mmBERT-base)

Licence

Code MIT; weights tagged MIT on Hugging Face

Run it

Load the backbone, head.safetensors and calibration.json using the Hugging Face card's Python example; the standard inference widget cannot load the separate head

Release date

First Hugging Face repository creation: 01 Nov 2026; v1.3 card checked 05 Nov 2026

Calibration

Optional label-specific additive biases can change the selected answer; temperature alone cannot

• For v1.3, the maker reports 0.823 micro accuracy and 0.0185 ECE on 13,373 decisions, with temperature fitted on a disjoint half. Its 0.0159 ECE fitted and scored on the same data is in-sample. vendor reportedsource
• On six tasks with identical rows and option wordings, v1.1 and v1.3 each get 8,120 of 10,157 decisions right (0.7994). The card withdraws its claim of better calibration after temperature fitting: split variation reverses the comparison. Word-sense questions supply 84% of these rows. vendor reportedsource
• On the fifteen routing categories whose jargon labels do not coincide with training wordings, the maker reports accuracy rising from 0.624 for v1.1 to 0.838 for v1.3. Each recipe was run once; this is not a multi-seed result. vendor reportedsource
• The maker reports v1.3 accuracy 0.568 rising to 0.640 on news rubrics under label-specific bias fitting; v1.1 reaches 0.673 under the same fit. The fitted biases belong to that option set and do not transfer to another set. This is a calibration-tool update, without a new backbone release. vendor reportedsource

Source: Weights and v1.3 model card → Source: GitHub → Source: Release metadata →

Send a shared state, named questions and optional images to these multimodal models. A single forward pass returns an answer distribution for every question, without generating text. The Qwen3.5-based family has 0.8B, 2B and 4B sizes.

Sizes

0.8B, 2B and 4B

Base

Qwen3.5; the 4B model uses Qwen3.5-4B

Licence

Apache-2.0

Inputs

Text and images

• InternLM reports a seven-benchmark average of 90.02 for Intern-Decision-4B against 88.74 for Jev, with Brier / ECE of 0.347 / 0.065 against 0.358 / 0.095. vendor reportedsource
• The model card reports mean latency of 44.16 ms for the 4B on one RTX 4090, against 109.70 ms for Jev. The local and network paths differ, so these timings are not like for like and depend on hardware and workload. vendor reportedsource

Source: 4B weights and model card → Source: GitHub → Source: Report → Source: Hugging Face upload metadata →

This 2B specialist scores whether a candidate OCR reading matches a document region. It handles text, formulas and tables, returning a score from 0 to 1. Qwen3.5-2B supplies the backbone; a binary classification head replaces its language head. Scores can help select among OCR engines or route uncertain regions to a person.

Size

2B

Base

Qwen3.5-2B

Licence

Apache-2.0

Access

Open weights in BF16, AWQ, GPTQ and GGUF; a free API on ModelBest's platform

• OpenBMB reports decision AUROC ×100 of 93.6 on its JevOCRBench, against 90.1 for a row named JEV-35B-A3B and 93.8 for Qwen3.5-397B-A17B. The card does not identify JEV-35B-A3B further. vendor reportedsource
• Across four OCR parsers on the same 200 samples with manually labelled readings, OpenBMB reports AUROC ×100 from 91.9 to 93.4. These are its own parser and sample combinations. vendor reportedsource

Source: Weights and model card → Source: Code and deployment → Source: JevOCRBench → Source: Hugging Face upload metadata →

Vectorize's Hindsight team fine-tuned this 0.8B model from Qwen3.5-0.8B for agent memory. It returns probabilities for choice, score and yes/no questions in one pass. Hindsight first tried these judgments with Jev, then trained its own model to check whether facts are new and evidence is enough. It uses the same System One request shape.

Size

0.8B

Base

Qwen/Qwen3.5-0.8B

Licence

Apache-2.0

Inputs

Text and structured state

Languages

English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese

Context

65,536 tokens per request (trained)

Price

$0.02 per million input tokens on Hindsight Inference; input only · Compare prices

Access

Open weights; vLLM plugin or Transformers on CPU, Apple Silicon and supported GPUs; hosted POST /v1/systemone

• In Hindsight's own 1.2.0 evaluation, quality rises for six of seven LLMs. glm-5.3-flash goes from 73.8 to 87.1 and deepseek-v4.1-flash from 81.9 to 85.1; gemini-3.1-flash-lite goes from 87.1 to 86.6. That model is also the judge, so its rows are self-graded. Costs fall for five LLMs. vendor reportedsource
• glm-5.3-flash with the decision model scores 87.1 at $7.72 per thousand cases, against 89.0 at $147 for gemini-3.8-flash alone. Costs use list prices and an equal mix of four memory operations. vendor reportedsource
• Median decision-model latency is 167 ms, with 350 ms at p90. The replaced LLM calls had median latencies of 1 to 7 seconds, depending on the model. vendor reportedsource
• Reflect costs 4.6x less with gemini-3.8-flash, while quality drops from 94.1 to 91.2. With gemini-3.1-flash-lite, reflect quality drops from 94.1 to 85.3. vendor reportedsource
• Against three other 0.8B models, its mean across eight benchmarks is 75.6, ahead of 74.5, 73.4 and 66.6. On the public Decision Index it trails StartLux-Decision 0.8B Instant, 36.1 to 53.8. vendor reportedsource
• Across its evaluation, Hindsight reports 3.1 LLM calls per case with the decision model, against 4.5 before. This is an evaluation-wide mean; individual memory operations and workloads differ. vendor reportedsource

Source: Launch and evaluation → Source: Weights and model card → Source: Hindsight code → Source: System evaluation →

Abacus.AI's 4B decision model answers a typed question about audio, an image or text with a probability per option in one pass. It is a Gemma 4 E4B-it fine-tune. The family includes a 744M audio model and a LiteRT build for phones. Each 4B request takes one audio clip up to 30 seconds or one image, with up to 20 options.

Size

4B effective, 8.6B stored; 744M audio tier

Base

google/gemma-4-E4B-it; google/embeddinggemma-2 for the audio tier

Licence

Apache-2.0

Inputs

Audio, images and text; 2 to 20 options

Formats

Transformers; LiteRT-LM; Core ML int8 for the audio tier

• On Abacus.AI's 6,671-item Earshot benchmark, the 4B scores 60.1 audio skill with ECE 0.069. Gemini 4 Argon scores 59.4 with ECE 0.123, and Gemini 3.8 Flash 50.5 with ECE 0.204. The 0.7B audio tier scores 53.0 with ECE 0.086. vendor reportedsource
• Using the 0.7B audio tier first and escalating below 0.9 matches the 4B's 60.1 audio skill at 130 ms mean latency instead of 210 ms in Abacus.AI's run. vendor reportedsource
• Abacus.AI reports 57.9 on public Decision Index images and 56.6 on text. These are its own runs, not leaderboard submissions. vendor reportedsource
• The 3.45 GB LiteRT build scores 59.4 audio skill in a PyTorch replica of its quantized weights. Abacus.AI measures about 0.6 seconds per text decision on an iPhone 14 Pro Max. vendor reportedsource

Source: Weights and model card → Source: 0.7B audio tier → Source: LiteRT build → Source: Maker blog announcement → Source: X announcement: 3 Nov 2026 →

REAX 1.3 runs two model parts on one GPU and averages their per-option probabilities 50/50. One blends a Bittensor subnet 92 miner fine-tune with KnowLine; the other blends a miner fine-tune from Mammoth-1's lineage with Decision 1.0's d3-ultra. Its interface accepts Jev-style requests. The hosted service at decisionmodels.io and the Hugging Face account decisionmodels belong to a separate business unrelated to decisionmodels.ai.

Licence

Apache-2.0 for the blended work, subject to preserved upstream licences and notices; REAX Engine is MIT

Weights

About 115 GB in FP8; reference hardware is one RTX PRO 6000 Blackwell 96 GB

Inputs

Text; noul, choice and score questions

Access

Open Hugging Face weights and REAX Engine reax13 mode

Local input cap

32,000 proxy tokens, calculated from compact request JSON characters divided by four, rounded down; larger requests are refused.

• REAX reports 73.71 on all 140,178 scoreable public Decision Index 0.3 requests. Counting the one request above the released engine's input cap as refused gives 73.70. These are internal measurements, not official leaderboard scores. vendor reportedsource
• On 3,478 public items answered by all eight models, REAX reports 69.5 against 68.8 for d3-ultra, within noise, and 67.4 for KnowLine. The second draw was added after a partial result was known. The public subset cannot establish an official Full-score ranking. vendor reportedsource
• On a uniform 1,000-call sample, REAX measures 207 ms median and 655 ms mean latency, one request at a time on one RTX PRO 6000. Long multi-question requests can take minutes. vendor reportedsource

Source: Weights, results and upstream notices → Source: REAX Engine → Source: Announcement →

A Korean-specialised 9B decision model based on Qwen3.5-9B, trained with int8 LoRA at rank 32. It reads a state and typed questions, then returns probabilities over the options in one forward pass. The first public release supports choice and yes/no questions through a Jev-compatible /v1/systemone server.

Size

9B

Base

Qwen/Qwen3.5-9B

Licence

Apache-2.0, declared for the model weights

Inputs

Text; choice questions with 2 to 26 options and yes/no questions

Language

Korean-specialised; the model card also tags English

Limits

State up to 6,000 characters; unsupported requests return HTTP 422

Access

Public model.safetensors weights and serving code

• ONYULTEA reports 0.866 accuracy on a five-task KoBEST conversion with 2,600 items, and 0.635 on 3,000 korean-jev-decision-index items. Its LocalLLaMA typed-decisions score of 0.818 is labelled fitted. vendor reportedsource
• On the public Decision Index 0.3 suite, ONYULTEA reports index 30.7, raw 44.8 and 168 ms p50 latency. It says training and serving used one 24 GB consumer GPU. vendor reportedsource
• The card reports KMMLU accuracy 45.4%, against a 23.4% zero-shot base result. Its 8 November history calls this a label-conversion correction to an earlier approximately 30% figure, not a new weight release. vendor reportedsource

Source: Weights, licence declaration and results → Source: Serving and evaluation code → Source: Weight file inventory →

Morph's hosted 27B decision model is based on Qwen3.5. It answers choice, yes/no and score questions about text, JSON or images, returning probabilities without generating text. The System One endpoint supports several questions about a shared state, with automatic prefix caching for repeated input.

Size

27B; Qwen3.5 base, according to Morph

Inputs

Text, JSON and images; PNG, JPEG and WebP as inline base64 data URLs

Context

65,536 tokens

Price

$0.042 per million uncached input tokens; $0.008 cached; output free · Compare prices

Images

Up to 16 per request, 8 MB each; remote image URLs are unsupported

Access

POST /v1/systemone on Morph; morph/reflex on OpenRouter

• Morph reports five production calls with a satellite frame, a blank image and repeat inputs. Repeated inputs reused 1,216 of 1,328 tokens (91.6%); the first satellite call took 890 ms and its repeat 220 ms. Switching to the blank image changed the decision and missed the cache. These are observed calls, not latency guarantees. vendor reportedsource

Source: Documentation → Source: OpenRouter listing → Source: Launch post →

A 0.8B decision model fine-tuned from Qwen3.5-0.8B. It reads a shared state and isolated question segments in one forward pass, returning probabilities for choice, yes/no and score questions. The release includes serving code and weights without a text-generation head. Its non-commercial licence limits how it can be used.

Size

0.8B; about 1.5 GB in bf16

Base

Qwen/Qwen3.5-0.8B

Licence

CC BY-NC 4.0; non-commercial

Inputs

Text and JSON state; choice, noul and score questions

Access

Public decoder and decision-row weights, Python runtime and Decision Index engine

• GyrTechnology reports 47.75 chance-corrected skill and 60.73 raw on the public Decision Index 0.3 suite, with all 140,178 requests answered. Median/mean/p80 latency is 18/54/30 ms on one RTX 5070 Ti, one request at a time. These are maker results; the earlier v0.1 scored 42.79. vendor reportedsource

Source: Current card and weights → Source: Release files →

An encoder-only decision model family built on Ettin, with 0.4B and 1B sizes. It returns typed answers alongside evidence spans, a relation to the state and an abstention estimate. It encodes and caches the state before seeing questions. Beta-3 adds non-decisive rule outcomes and a field-extraction adapter; the first release was 20 Oct 2026.

Size

0.4B and 1B

Base

Ettin encoders from jhu-clsp

Licence

Weights, data recipe and docs: CC BY-NC-SA 4.0; code: MIT

Inputs

English text or JSON state; up to 2,048 state tokens

Access

Public weights; Python runtime; /v1/systemone, /v1/clairic and combined /v1/horopter endpoints

Training restriction

Training includes TypeSafe Jev API outputs. The card says TypeSafe terms prohibit using them to imitate its model or build a competing service.

• The maker reports ShARC four-way accuracy of 82.3 for the 1B beta-3 release, against 54.3 for its first release. ShARC training data is included, so the dev set is held out within the same distribution. On JevBench hard, the 1B scores 40.5 against 64.0 for clef-flash. Generated-policy results do not establish the same performance on real rule text. vendor reportedsource

Source: 1B model card and weights → Source: 0.4B model card and weights → Source: Runtime and recipes →

A 150M pull-request decision model built on TS-Ettin-150M. It asks five fixed questions about a diff, covering description agreement, review difficulty, compatibility, test coverage and security. Its 8-bit ONNX package runs the questions in one CPU pass. The questions use choice and yes/no answers; this is a specialist package.

Size

150M

Base

typestrict/TS-Ettin-150M

Licence

Apache-2.0 tag on the model card

Inputs

English pull-request descriptions and code changes

Access

164,505,784-byte ONNX file, tokenizer and decision_ai.json; ONNX Runtime or decision_ai for Flutter

• The maker reports wins on four of five questions under its refund calculation, which combines four metrics and bootstrap intervals. On description agreement, macro-F1 is 0.928 against Jev 0.807 over 436 held-out examples. On compatibility, Jev is ahead: 0.910 against 0.870 over 2,000 examples. These are task-specific vendor results. vendor reportedsource

Source: Model card and package → Source: Model page and audit → Source: Release history →

A d3-flash fine-tune with typed choice, yes/no and score answers for text, JSON, images and videos. It keeps the Qwen3.5-9B multimodal architecture and decision readout. The bundled Transformers runtime extends score questions to 255 ordered levels; the documented vLLM build keeps the upstream 10-level limit.

Size

8.394B, including a 0.456B vision encoder

Licence

Apache-2.0; LICENSE and NOTICE shipped

Inputs

Text, JSON, images and videos

Context

262,144 configured positions; this is not a GPU memory guarantee

Access

Full BF16 weights and FP32 readout; Transformers or /v1/systemone on the pinned vLLM nightly

• The card reports a validated deployment on one RTX 5090 with a pinned vLLM nightly and recorded dependencies. Each question gets its own forward pass, batched by the runtime. It publishes no task-accuracy benchmark. vendor reportedsource

Source: Model card and weights → Source: Release history →

A pre-release 4B d3-mini fine-tune for reviewing code written or run by an agent. It reads the task, repository baseline, behaviour summary and change, then returns three fixed decisions about class, intent and scope. The static extractor never runs the code being inspected. It is intended to rank items for further review.

Size

4B

Licence

Apache-2.0; LICENSE and NOTICE shipped

Inputs

Text task and diff or executed snippet; truncated above 8,192 tokens

Access

Merged weights, readout, adapter, extractor, frozen questions and thresholds

• At its frozen thresholds, the maker reports 458 of 510 positives found across 14 languages and shells, with 11 of 2,150 real items and 1 of 560 authored benign controls flagged. On an independently written agent-diff set it finds 21 of 22 positives with 0 of 172 negatives flagged. These evaluations were run by the maker. vendor reportedsource
• Training positives came from one writer model; recall falls on other writers. Unrequested personal data in request headers is mostly missed (1 of 40). The release is a pre-release and is not a sole security gate. vendor reportedsource

Source: Model card, weights and results → Source: Release history →

A public-preview family for simulated survey respondents, fine-tuned from Kev at 4B and 27B. It returns option probabilities from persona traits and a question, with training intended to make answers respond to prices, odds and payoffs. The 27B release has merged weights; 4B has a LoRA adapter and pointer head. Its tests check economic consistency, with no validation against real people.

Size

4B and 27B

Licence

Apache-2.0 tags on both model cards

Inputs

English personas and survey questions, using US-style amounts and settings

Access

kev.serve at the pinned Kev revision; /v1/systemone; EDSL survey adapter included

• The maker reports Lory-27B scoring 88 on its own nine-family DLS test against Kev-27B 71, and 92 against 90 on 513 contrasts written separately by GPT-6.1-sol. The blind test measures direction of movement, not its size. Each size has one training seed. vendor reportedsource
• Lory-4B scores 79 on DLS and 81 on the blind test, against released Kev-4B 53 and 67. Its framing-invariance score is 56/100, against 66/100 for Lory-27B. Economic consistency does not establish realistic human responses. vendor reportedsource

Source: 27B card and merged weights → Source: 4B card and adapter →

A visual-inspection decision model built around Muse Glimmer 30B. The release supplies two small logistic heads for a fixed manufacturing-defect question, plus a temperature for zero-shot scoring; the backbone is downloaded separately. A head must stay paired with its exact prompt, image resize and object noun.

Size

30B backbone; each logistic head file is 134,696 bytes

Licence

Heads and code Apache-2.0; MVTec AD data CC BY-NC-SA 4.0, research use only

Inputs

English questions and part images resized to a 448-pixel long side

Access

Public heads and calibration files; Muse Glimmer 30B required separately; code and demo on GitHub

• The maker reports AUROC 0.962 and 90.2% accuracy on the held-out 866-image half of the same 15 MVTec AD products used to train the head. The layer-13 head reports 0.954 and 88.3%. These are known-product tests, not general defect-detection accuracy. vendor reportedsource
• The card reports 45 ms per image batched on one H100 for the full model and 21 ms for a truncated layer-13 benchmark. The supplied server still runs the full depth, so loading the early head alone does not give the 21 ms path. vendor reportedsource

Source: Heads and model card → Source: Code and run records →

A multilingual Laya fine-tune that chooses a proposed value for a dialogue slot or answers not specified. It needs candidate values from another component and does not extract spans. The released seed was selected on development data from three training seeds; its card reports results for 46 training languages and five held-out languages.

Size

About 322M, according to the card

Licence

CC BY-SA 4.0 weights

Inputs

Dialogue message, context and candidate values; 1,024 input tokens and 48 per option

Access

Laya runtime; public model.safetensors, configuration and calibration files

• On MTOP across six languages, the authors report slot F1 0.953 with oracle candidate lists and 0.488 with real proposers, whose candidate recall is 0.578. With real lists on MTOP and MASSIVE, ECE remains about 0.15 to 0.17. Thresholds and temperature need fitting to the user's own data. vendor reportedsource
• At the card's 1% target risk, the development-set threshold keeps 60% of correct fills with oracle lists but only 3% with real lists. These are development-set operating points, not guaranteed test error rates. vendor reportedsource

Source: Model card and weights → Source: Code →

Finic's hosted fraud and risk decision model returns probabilities for yes/no, single-choice and multiple-choice questions. The ask endpoint uses grayson-1 and accepts structured case data or free-text context. The separate score endpoint builds a 0 to 1,000 risk score from a scoring model calibrated with labelled records. An HN comment surfaced it on 04 Nov 2026; its release date is unconfirmed.

Inputs

Structured data and free-text context

Context

32,768 input tokens per request

Questions

1 to 128 per ask request; choice and multi_choice have 2 to 255 options

Price

$0.035 per million input tokens for ask and score; creating and calibrating scoring models is free · Compare prices

Access

Hosted POST /v1/ask with model grayson-1; no public weights confirmed

• Finic documents identical answers for an identical request, while changes to wording, option order, data, context or key order make a different request. It also says adding or changing another question can move numerical values slightly, under one percentage point on average. This is the vendor's contract, not an independently reproduced determinism test. vendor reportedsource

Source: Vendor and model identity → Source: Typed questions → Source: Pricing → Source: API contract →

A Qwen3.5-35B-A3B decision fine-tune with 3B active parameters. It reads text, chat or up to four images and answers typed questions from option-letter logits, without generating an answer. BF16 and NVFP4 packages are public under Apache-2.0; the maker identifies NVFP4 as its production format.

Size

35B total parameters, 3B active

Inputs

Text, chat, up to 4 images

Limits

131,072-token context; up to 64 questions and 255 choice options

Licence

Apache-2.0

• The maker reports 71.33 on the public JDI 0.4.1 suite from its own run. This is not the full index score or an independent reproduction. Its 41 ms median latency has no hardware or workload specification in the card. vendor reportedsource

Source: BF16 card and weights → Source: NVFP4 package → Source: Release history →

Qwen3.5 typed-decision models in 9B v5 and 0.8B v5 packages, with a separate 255-option head and noncausal attention in their full-attention layers. Both include merged weights and the Decider-derived System One server. Overlong question branches are refused instead of truncated.

Sizes

9B v5 and 0.8B v5; BF16 decision heads

Context

8,192 tokens per question branch

Licence

Apache-2.0; serving code credited in NOTICE

• The maker reports public Decision Index 0.3 skill 69.28 and raw 76.81 for 9B v5 over 140,620 requests. All 345 overlength HTTP 422 refusals count as wrong. The run uses BF16 CUDA on one A800 80 GB per server process. vendor reportedsource
• For 0.8B v5, the maker reports skill 54.69 and raw 66.33 on 140,620 public-suite requests, including 345 refusals counted wrong, using BF16 on one A800 80 GB per server process. The preceding v4 card reported skill 55.10; this is a lower score. vendor reportedsource

Source: 9B v5 card and package → Source: 0.8B v5 card and package →

A Qwen3.5-0.8B model with seven fixed heads for Chinese enterprise group chats. One pass labels category, sentiment, importance, urgency, intensity, attention and to-dos. K-IM-Deiction V2 ships merged weights, an optional adapter and an MLX port. The private chat data is not released.

Package

Merged Qwen3.5-0.8B V2 backbone, seven heads and calibration files; optional rank-64 adapter and native MLX edition

Inputs

Text windows, left-truncated to 512 tokens

Licence

Apache-2.0

Newer package

V2 weights uploaded 6 Nov 2026; MLX edition 7 Nov 2026

• On 1,000 windows from seven unseen groups, the maker reports category macro-F1 0.611 across seven classes, or 0.673 after excluding classes with support below ten. Needs-attention AUC is 0.822; binary-head ECE is 0.053 and 0.069. Labels come from GLM-5.3, with a second teacher used for agreement checks. Median latency is 41.0 ms on an RTX 4090 D at batch one. The earlier DeepSeek comparison was not rerun for V2. vendor reportedsource

Source: K-IM-Deiction card and merged weights → Source: Package history →

Two Kev-4B adapters check whether linked authority records describe the same place and whether a declared parent contains a child place. They supply pointer heads and calibration temperatures, with the Qwen3.5-4B base downloaded separately. The part-of adapter continues training from the equivalence adapter.

Inputs

Rendered authority records and fixed yes/no statements

Package

LoRA adapters and pointer heads; separate 4B base

Licence

Apache-2.0 tags on both cards

• The equivalence maker reports balanced accuracy 0.940 on 499 held-out decisions over 150 places. Labels retain agreement between Gemini and Jev and were not checked by people, so this measures agreement with machine consensus. vendor reportedsource
• The part-of card reports balanced accuracy 0.983 and Brier 0.014. Only six real negative links appear in its test split; synthetic negatives and authority hierarchy evidence limit conclusions about rare real errors. vendor reportedsource

Source: Equivalence card and adapter → Source: Part-of card and adapter →

An independent Qwen3-0.6B text decision starter with a frozen backbone and a trained 1,025-parameter compatibility head. It uses no LiquidAI weights. Each question-option pair repeats the state in its own input row, so candidate count adds work and memory. The package includes weights, code, splits and a training report.

Inputs

Text or JSON; choice, independent multiple choice, yes/no and score

Context

512 tokens per candidate row; longer input raises an error

Licence

Apache-2.0

• The saved report gives trained accuracy 0.5715 versus baseline 0.386 on 2,000 test decisions across Banking77, SST5, GoEmotions, BoolQ and AG News. The checkpoint and temperature were selected on a separate calibration split. These task results do not validate arbitrary questions or other domains. vendor reportedsource

Source: Card and standalone package → Source: Training report →

A Qwen3.5-4B-Base LoRA and Kev head for authorising outgoing tool calls. Bootstrap code fetches the pinned backbone and Kev source. The card leads with the full held-out test and warns that a threshold calibrated on development data transfers badly across domains.

Package

LoRA adapter, pointer head, inference code and evaluation records

Licence

SecJev Use and Distribution Agreement v1.0; custom terms

• The maker reports AUROC 0.8661 on 7,192 full-test rows. At the development threshold 0.0021, test true-positive rate is 0.776 and false-positive rate 0.221, versus development false-positive rate 0.010. No independent replication under its frozen chain is reported. vendor reportedsource

Source: Card, licence and records →

A local CodeSync bundle with 29 fixed judgments about user turns and shell commands. The same v3 release now includes Q8_0 and Q4_K_M GGUF files plus a standalone Python runner for Ollama. Unsupported questions inside CodeSync go to its cloud judgment service. No model licence is supplied.

Questions

21 turn judgments and 8 shell-command judgments

Access

CodeSync bundle; standalone Ollama runner and GGUF files

Licence

Not specified in the card or file inventory

Runtime

Ollama makes one request per question; the built-in runtime answers a turn in one forward pass

• DevsHub reports 97.9% on 233 hand-labelled decisions, 95.5% on 3,267 held-out session decisions and 93.7% on 4,120 rare cases. Its 13 ms median is measured on an unspecified laptop GPU. The card does not establish general question-answering performance. vendor reportedsource
• The maker reports Ollama Q8_0 accuracy 97.9%, 95.7% and 93.6% on the 233, 3,267 and 4,120-decision sets, with about 80 ms median per eight-question command on a laptop GPU. These are serving-format checks of v3, not a newly established trained checkpoint. The runner only classifies the supplied command; it never executes it. vendor reportedsource

Source: Card and bundle → Source: Standalone runner →

A Qwen3.5-2B LoRA for one forum-spam question. It reads yes/no token logits as a spam probability after supervised training on the dataset’s training split. It is independent of TypeSafe’s hosted Jev and needs the separate Qwen base.

Task

Binary forum-spam detection; text truncated to 1,500 characters in evaluation

Package

PEFT LoRA adapter; separate Qwen3.5-2B base

Licence

Apache-2.0

• The maker reports 2,723 of 2,725 test rows correct, F1 0.9993 and ECE 0.001, with two false positives and no false negatives. Spam examples are largely synthetic promotions. The fine-tune saw this dataset’s training domain while the compared open models were general-purpose; no result against TypeSafe’s Jev is claimed. vendor reportedsource

Source: Card and adapter →

A text-only d1-2B fine-tune that decides whether transcribed speech is for a voice assistant, still unfinished or meant for someone else. It runs through MLX on Apple Silicon. The fixed question was trained around the names Toby and Heimdall; the state can supply other names. The weights preceded this scan window; the in-window card explains their lineage.

Inputs

Conversation transcripts; speech recognition runs first

Formats

BF16, int8 and int4 packages

Licence

LFM Open License v1.0; commercial revenue limit applies

• The maker reports two training stages totalling about 175 million tokens across ten languages. The card gives no held-out dispatch accuracy, so training scale does not establish reliable wake-word-free behaviour. vendor reportedsource

Source: Card, weights and MLX runner →

A European hosted decision-model family with 9B and 27B API choices. It accepts text plus yes/no, choice and scale questions through /v1/zuse, returning probabilities without generated answers. The paper studies 4B, 9B and 27B adapters; the public API page lists 9B and 27B. Its release date is unconfirmed.

Hosted models

zuse-9b and zuse-27b

Inputs

Text; yes/no, choice and scales; up to 255 choice options

Context

65,536 tokens per document on the API

Price

$0.12 per million input tokens for 9B; $0.75 for 27B; no output tokens · Compare prices

Access

Bineric API; dedicated and on-premises deployments by arrangement

• The author reports 87.5% accuracy for 27B on a 27-language development suite that also guided model selection, with one seed. Pooled ECE is 0.016 and source-weighted ECE 0.035. These are development results, not untouched test accuracy. vendor reportedsource

Source: API, limits and prices → Source: Paper and evaluation limits →

Qwen3.5-derived decision models with text exits at different depths. The 4B package has 32 decoder layers; the 3.30B sibling keeps its first 20 layers and existing heads without retraining. Both read text and up to four images, returning option probabilities without generated text. Images use the full depth of the chosen package. Their runtime supplies option echo and prefix-bidirectional attention.

Size

4B package: 4.69B parameters and about 9.7 GB; 3B package: 3.30B and 6.8 GB

Inputs

Text and up to four images; choice, noul and score

Licence

Apache-2.0

Limits

32,768 tokens; up to 5,120 options; overlength requests get HTTP 422

Access

Public weights with the RSI-Jev runtime; no causal prefix cache

Newer package

3B (cut at layer 20) uploaded 7 Nov 2026

• The maker reports 60.61 on the public Decision Index 0.3, with 29.4 ms median and 236 ms p95 on one RTX PRO 6000 Blackwell at default effort. This is a public-suite run, not the full private index. Training used one seed and unreleased data. Ambiguous KoBBQ questions and some medicine and customer-support tasks regress against v6.1-VL. vendor reportedsource
• For the released 3B package, the maker reports public Decision Index 0.3 score 60.01, 19.6 ms median and 151 ms p95 on one RTX PRO 6000 Blackwell. All 4,001 checked rows match the 4B at medium effort bitwise. Against the full-depth 4B, held-out accuracy falls from 0.811 to 0.797 and MMLU-Pro from 0.502 to 0.469. The package reuses the 4B temperatures and thresholds. vendor reportedsource

Source: Model card and files → Source: 3B package and correction to the 4B release score →

A Korean-focused Laya multilingual fine-tune. Release 2026.11.08.1 continues from the previous original float checkpoint on private knowledge-graph and research summaries. It ships armD-final-epoch-3 as Q8_0 and original float weights. The maker records a release override with prior-task retention and news checks still unrun; it is not a selection-protocol pass.

Inputs

Text; Korean and English

Licence

Apache-2.0

Access

Self-contained macOS arm64, Linux x86_64 and Windows x86_64 runtime bundles; original float checkpoint for laya 0.4.22. The Laya GGUF is not an Ollama or llama.cpp chat-model package

• The maker reports 155/178 calibration and 153/185 test summaries correct, versus 135/178 and 132/185 for the previous release. Labels come from two Opus 5.5 passes without human adjudication. Calibration selects the epoch and temperature; the test split compared all four experiment arms, so no independent final test is claimed. Existing news cohorts, judgment-retention tasks and warm-daemon latency were not measured for this release. vendor reportedsource

Source: Model card and files →

Four local move-ranking fine-tunes for Spot’s Biscuit Snake: d1-2B, d1-3B, MiniMind-JEV and Laya Multilingual. The game supplies safe moves; the model ranks them to imitate a heuristic teacher that leaves the most open room. Keep the game’s collision checks. The d1-2B, d1-3B and MiniMind-JEV weights were uploaded on 5 November; the Laya Multilingual weights were uploaded on 6 November.

Variants

d1-2B and d1-3B merged weights with GGUF formats; MiniMind-JEV ranking head and BGE encoder; Laya Multilingual fine-tune

Inputs

Game state and candidate safe moves

Licence

d1 variants: LFM Open License v1.0 plus Apache-2.0 contributions (LvW-license-Snake-1.0); MiniMind-JEV: MIT; Laya Multilingual: Apache-2.0

Calibration

Laya Multilingual: temperature fitted on 400 reserved training-split rows, returning approximately 1.0. d1-2B, d1-3B and MiniMind-JEV: confidence was not recalibrated after training

Multilingual release

Laya Multilingual fine-tune; English and Dutch; laya 0.4.26

• On 1,708 held-out moves, the maker reports 99.82% teacher agreement for d1-3B and 55.80% for MiniMind-JEV. Saved-file reload checks cover 32 rows. Teacher agreement measures imitation of the heuristic; game scores, win rates and speed were not measured. vendor reportedsource
• The multilingual card reports 99.94% teacher agreement on 1,600 challenge cases, with English and Dutch using the same boards. It reports ten-bin ECE 0.0483. These are teacher-label checks; game score and win rate need separate rollouts. vendor reportedsource

Source: Model card and files → Source: d1-2B card and weights → Source: MiniMind-JEV card and weights → Source: Multilingual card and weights →

A LoRA decision adapter on LLaDA-MoE-7B-A1B-Instruct. One masked forward pass returns a probability distribution over option letters, with no generated answer. The base weights must be downloaded separately. Use its bundled inference script for the recorded scores; it repairs a rotary-embedding compatibility issue.

Base

LLaDA-MoE-7B-A1B-Instruct; separate base download

Licence

Apache-2.0 adapter, code and base

Limits

2–26 letter-labelled options, six recommended; 4,096 tokens recommended, up to 8,192 supported

Access

Public LoRA adapter and inference.py

• The maker reports 195 of 231 public JevBench v1.2.2 tasks correct (84.4%), versus 155 for the base using the same script. Published competitor rows use different serving protocols. On 66 evidence-silent VitaminC cases, it selects “not enough info” for only 16. Calibration fitted on English decision data needs checking on other tasks. vendor reportedsource

Source: Model card and files →

A decision adapter and readout head trained on the text decoder of Nemotron-3-Nano-Omni-30B-A3B. The adapter combines three training runs. The same weights can read images, video and audio through the base’s frozen encoders using the tagged multimodal runtime. Media were not used in training. The repository’s main branch and released package versions cannot load this base.

Base

Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 text decoder

Inputs

Text; zero-shot images, audio and video with at most 16 frames; choice, noul and score

Licence

NVIDIA Open Model Agreement for the model derivative; code Apache-2.0

Hardware

62.5 GB bf16 base download; one 80 GB GPU or multiple GPUs, plus CUDA Mamba kernels

Access

Public adapter and head; swiftpick-30B-A3B-v1-mm runtime tag

• The maker reports 183 of 231 public JevBench items correct (79.22%), Brier 0.287 and ECE 0.059 at a 4,096-token window. Its separate 42.4 Intelligence estimate uses only the open half of the v1.5 method and omits sealed items and other axes; it is not an official score or rank. Questions are sometimes under-read, and calibration transfers poorly beyond the fitted short tasks. vendor reportedsource
• In one H100 bf16 serving evaluation, the maker reports NaturalBench accuracy 0.820 on 1,200 items, MINDS-14 en-US 0.986 on 563 and Perception Test video 0.720 on 300. Removing media reduces performance, but mean confidence remains 0.74 to 0.82 on image and keyword sets. Media presence needs checking; these few public sets do not establish broad multimodal reliability. vendor reportedsource

Source: Model card and files → Source: Multimodal runtime and evaluation →

A Qwen3.5-4B full decision fine-tune that returns probabilities for typed questions in one forward pass. The v2 release replaces the earlier full-model files with a decision-only package; v1 remains tagged. Its licence restricts use to non-commercial purposes because some training sources are research-only.

Base

Qwen/Qwen3.5-4B

Inputs

Text state; choice, noul and score

Licence

CC BY-NC 4.0; non-commercial

Access

Public decision weights, runtime and Decision Index engine

• GyrTechnology reports 57.71 chance-corrected skill and 68.61 raw on the complete public Decision Index 0.3 suite, with no unsupported requests. This is the maker’s run, not an independently reproduced full-index rank. vendor reportedsource

Source: Model card and files →

A text-only Gemma-4-31B-it fine-tune with a linear readout over supplied options. One pass answers choice, noul and score questions. The checkpoint contains vision weights, but the runtime does not use them. Its probabilities have not been calibrated.

Size

31.3B bf16 parameters plus a 255 by 5,376 fp32 readout

Inputs

Text or JSON; up to 255 options

Context

32,768 tokens per sequence; multi-question requests can be split, with the full state in each part

Hardware

62.5 GB of weights; longer requests need additional memory

Licence

MachinePulse Community License effective 12 Nov 2026. Earlier Apache-2.0 copies retain their rights; inference code is separately Apache-2.0

Commercial terms

The weight licence effective 12 Nov 2026 requires attribution for external commercial use and a separate agreement for commercial production above US$10M annual revenue, including affiliates

• The maker reports public Decision Index 0.3 score 75.05 across 140,178 scored requests, and private Full score 73.86 on the 11 Nov 2026 Decision Index 0.4.1 board. The private result is a maker report, not a reproduced test. vendor reportedsource

Source: Card and weights → Source: Weight licence and effective date → Source: Separately licensed inference code →

An independently maintained 2.21B typed-decision model with a frozen vision tower. Its diagnostic continuation raises the reported public score but fails its pre-registered calibration guard. The required renderer, attention mask and verdict readout ship with the weights; generic chat runtimes do not reproduce its probabilities.

Size

2,213,241,664 parameters including the vision tower

Inputs

Text and images; image smoke checks, without a new full vision benchmark

Context

32,768 tokens; up to 255 options; overlength inputs refused

Licence

Apache-2.0; upstream data and teacher-output terms still apply

Access

Merged weights and custom runtime; separate text-only GGUF and MLX packages

• The maker reports 46.74 on the complete public Decision Index 0.3 suite, versus 44.37 for its parent, with 140,178 successful requests. ECE rises from 0.0195163271 to 0.0300812254, exceeding the allowed 0.01 increase. Concurrent CUDA timings are not serial request latency. vendor reportedsource

Source: Card and weights →

Two multilingual Laya fine-tunes copy Jev judgments for Telegram messages and Farcaster casts or accounts. The Telegram bot uses a 0.81 deletion gate and exempts admins. Farcaster ships Platt scaling for its raw scores. The evaluations measure agreement with the teacher, including its mistakes.

Base

convaiinnovations/laya-multilingual

Inputs

Text, links and metadata; no media-content or avatar inspection

Training

Telegram: 5,540 messages, with training data unreleased; Farcaster: 16.6k casts and 4,384 account snapshots

Licence

Apache-2.0, according to both model cards

Access

Public weights, question files and loading examples for both classifiers

• The maker reports 95.4% agreement with Jev on 978 held-out Telegram messages at the 0.81 gate, precision 0.82 and recall 0.70. These compare with Jev's delete labels, not human judgments. Reported median latency is 38 ms on an Apple GPU and about 400 ms on a few CPU threads. vendor reportedsource
• The maker reports 95.7% agreement on 2,489 held-out Farcaster casts at gate 0.5, precision 0.92 and recall 0.80. Account agreement is 97.1% on 652 held-out accounts. The newest 15% is held out by time; new spam styles may differ. vendor reportedsource

Source: Telegram card and weights → Source: Farcaster card and weights →

A Gemma-4-31B-it decision family served through Flide. The 1.5g package describes screenshots, photographs and text; the newer 1.6g card documents text and choice, noul and score questions. No new image capability is inferred from the earlier package, and neither card supplies a measured accuracy or calibration result.

Base

google/gemma-4-31B-it

Inputs

1.6g: text. The earlier 1.5g card describes text and images

Licence

Apache-2.0, according to the card

Access

Merged safetensors and a Flide serving example

Source: Card and weights → Source: 1.6g card and weights →

Poindexter Labs publishes 31B and 57B structured-decision packages with System One servers. The 57B combines Gemma-based and Qwen-based components and accepts text or JSON. The 31B Transformers path reads text; its image path requires compatible SGLang serving. Probabilities need validation on the deployment task.

Size

Approximately 31B, or 57B across two components

Inputs

57B: text and JSON only; 31B image support depends on a compatible SGLang backend

Questions

Up to 64 per request in the reference server

Licence

Apache-2.0, according to the card

Access

Public 31B and 57B weights with separate reference serving code

57B limits

64 questions per request; 2 to 255 choices; 40,000-token input proxy and 32,768 tokens per rendered input; overlength requests refused

Newer package

57B weights uploaded 6 Nov 2026

• Poindexter Labs reports 75.34 on the full public Decision Index 0.3 suite of 140,178 requests for 57B. Its latency sample covers only 300 public-suite requests, with 270 ms median on one RTX PRO 6000. These are maker measurements, separate from the index maintainers' private evaluation. The reference setup uses a 96 GB NVIDIA GPU and about 110 GB of model files. vendor reportedsource

Source: Card, code and weights → Source: 57B card, weights and serving requirements →

A research d1-2B fine-tune for proposed agent actions, followed by a 10,240-weight output-head edit. Fixed questions cover risk, review, injection and authorization. The card corrects an earlier evaluation whose question selection revealed labels. Long states are truncated at 8K tokens, and authorization remains weak.

Size

1.92B parameters with an untied output head

Inputs

JSON action, user request and surrounding context; fixed trace or command question sets

Context

128K architecture context; states trained and evaluated only up to 8K tokens

Licence

LFM Open License v1.0

Access

Weights, custom Transformers code, calibration, questions and edit provenance

• The maker reports 0.971 AUROC on 2,808 LlamaFirewall traces and 0.736 AUROC for review on 822 ATBench traces. Authorization balanced accuracy is 57.6% on 621 AuthorizationBench items and 32.5% on 175 Mailu actions across three classes. All traces receive the same three questions; 451 LlamaFirewall traces are truncated. These results do not establish a production safety control. vendor reportedsource

Source: Card, weights and provenance →

A Sarvam-105B decision fine-tune, offered as custom FP8 weights and a hosted API. It returns option probabilities without generating text. Choice options are sorted by key before scoring. The hosted API and downloadable runtime have different context limits.

Size

105B; about 106 GB of weight files

Inputs

Text or JSON; no images

Context

Hosted: 32,768 tokens per question; local runtime: 131,072

Hardware

Local FP8 runtime needs about 120 GB free GPU memory; tested on B200

Licence

Apache-2.0

Price

Sarvam API: ₹4 per million input tokens; output free · Compare prices

Access

Public weights and prism.py; hosted POST /v2/systemone

• Sarvam reports 44 ms median and 420 ms p95 on 1,000 serial public-index requests on a B200. vendor reportedsource
• In its 2,000-question shuffle test, fixed-order preprocessing yields zero changed answers; scoring options as sent changes 4.0%. Sorting is part of the tested system. vendor reportedsource

Source: Launch and evaluation → Source: Card, runtime and weights → Source: Hosted API limits and price →

A text/image decision model built on Kanana-2-v-3b. It compares state/question embeddings with candidate embeddings and returns option probabilities. Candidate embeddings can be cached. The supplied HTTP server processes one request at a time.

Size

3.67B: 3.18B language model and 0.49B vision encoder

Inputs

Korean or English text, JSON or chat; up to 10 images; no video

Context

20,480 tokens; overlong inputs are shortened by default, or rejected with --strict-inputs

Licence

Kanana Open License

Access

Public weights, custom Transformers code and /v1/systemone server

• Kakao reports 61.34 on public Decision Index text and 61.12 on vision, averaging 61.23 across those two scores. Its Korean benchmark score is 72.50 over 15,909 questions. These are maker results, not a combined ranking of all decision models. vendor reportedsource

Source: Model card and weights →

A Qwen3.5-2B fine-tune with an Unsloth Clef decision head, intended to choose sections in table-of-contents retrieval. It was trained on general typed decisions. Retrieval, citations and final answer writing remain the application’s job; table-of-contents routing quality is unmeasured.

Base and size

Qwen3.5-2B; merged backbone plus separate decision-head weights

Inputs

Typed state and choice, noul or score questions

Context

Decision config records max_len 4,096; backbone settings are not a tested task limit

Licence

Apache-2.0

Access

Public weights; Unsloth FastDecisionModel required for typed inference; hosted handler deployment untested

• The training record reports 66.0% accuracy and 0.0948 ECE over 150 held-out decisions, against 37.33% and 0.0967 for the base. The split, seed and evaluation script are absent. These are general decision results, not a retrieval benchmark. vendor reportedsource

Source: Model card and weights →

A Kev-0.8B fine-tune that judges shell commands for termite. It asks whether a command only reads, keeps secrets hidden or could do harm, and can rank candidates. The 0.2 weights are rebuilt from documented training sources. In termite its judgment can only tighten the rules; harmful commands can still pass the model.

Size

0.8B; 812 MB Q8_0 GGUF

Inputs

English shell commands

Runtime

llama.cpp with /v1/systemone; only that server was tested

Licence

Apache-2.0; 0.2 uses CC BY 4.0 tldr-pages text and project-written rows

Access

Public GGUF weights

• On the maker’s fixed 141-command allowed test, 0.2 rates 5 of 48 harmful commands safe and 46 of 68 harmless reads safe. Its risk-order tau is 0.678, down from 0.698, and the safest candidate ranks first in 3 of 10 ladder choices. The ladder informed run selection and three acceptance marks were set after initial scores were seen. vendor reportedsource
• The model does not check whether a command exists. The card says rules catch missing programs and cut-off lines, but not made-up options. Linux was not tested separately for 0.2, and the unpublished rule-labelled training rows do not establish independent safety correctness. vendor reportedsource

Source: Model card and weights →

A Qwen3.5-4B LoRA and Unsloth decision head for next-action and tool routing. Training used 17,600 source-labelled rows; the name refers to the larger merged data pool. Jev answers were excluded from training. It is unsuitable as a general trace judge.

Base and size

Qwen3.5-4B; adapter and separate decision head

Inputs

Software-agent and customer-service traces

Context

4,096 tokens; reported test states up to 13,000 characters

Licence

Apache-2.0

Access

Public adapter; Unsloth FastDecisionModel and base weights required

• The maker reports 86.5% agreement with source labels on 480 held-out rows, against stored Jev 1.13 answers at 80.6%. These labels describe the source trace’s next action or parsed tool call, not human correctness. Trace-disposition results are weak. vendor reportedsource

Source: Model card and weights →

A Gemma 4 12B fine-tune with merged weights for the decisio runtime. Its commit rule pushes yes/no probabilities inside the 0.20 to 0.80 band to that band’s edge on the same side. This preserves the winning answer while changing calibration.

Base and size

Gemma 4 12B; merged bf16 weights

Inputs

State and typed choice, yes/no or score questions

Runtime

decisio 0.35.0 or later; plain vLLM does not apply the full readout

Licence

Apache-2.0, per card and included licence

Access

Public weight shards and decision_config.json

• The maker says temperatures were fitted by log loss and training was checked against both boards’ public test sets. Training data are undisclosed; the card supplies no task-accuracy result. vendor reportedsource

Source: Model card and weights →

Open 0.8B and 4B decision models that steer a separate chat model using rules supplied per request. Turn and partial-draft judgments can request tools, enforce an opening phrase or stop a reply. Rein returns decisions; the chat model writes the text.

Variants

Rein-r1-0.8B and Rein-r1-4B

Inputs

Rule sheets and text conversations; English and Portuguese best measured

Runtime

Custom Transformers decision and session-control code

Licence

Apache-2.0, per both cards

Access

Public merged backbone and decision-head weights

• The maker reports 96.8% decision accuracy and 0.003 ECE for 4B on 28,430 questions about held-out rules. Labels are synthetic and audited by another model. End-to-end chat tests still contain rule violations; they are small sampled sessions judged by Qwen. vendor reportedsource

Source: Model card and weights → Source: 0.8B card and weights →

A Qwen3.5-4B text/image decision fine-tune with merged weights and a Jev-like API. Its runtime uses bidirectional attention within the state and repeats some inputs. Plain causal inference produces different probabilities. This research preview is noncommercial.

Size

4B; merged weights

Inputs

English or Chinese text, JSON and images

Runtime

vev serve on an NVIDIA GPU; checkpoint-specific attention and repetition settings

Licence

CC BY-NC 4.0; Apache-2.0 base licence retained

Access

Public merged weights; no separate adapter

• The maker reports public vision skill 72.90 and text skill 60.32, with disclosed training/evaluation overlaps and a text-suite hash mismatch outside the index. Across 18 release sets, ECE ranges from 0.01 to 0.24; four sets regress versus Vev 4B. No full private-board score is supplied. vendor reportedsource

Source: Model card and weights →

A Qwen3.5-0.8B LoRA plus an Unsloth Clef head for fixed sheets of regulatory and general questions. The calibration split was left unused, so this release is uncalibrated. It is intended for triage and routing, with mixed expert and teacher label provenance.

Size

0.8B base with LoRA and separate joint decision head

Inputs

Text/JSON cases and fixed typed-question sheets

Context

4,096-token training setting; serving reads up to 16,384 without a matched long-context test

Licence

Apache-2.0 model tag; derived dataset CC BY 4.0

Access

Public adapter and head; Unsloth required, standard PEFT alone omits the head

• The maker reports 0.8906 accuracy on in-distribution legal questions and 0.7778 on the separate AIReg transfer set, with ECE 0.0237 and 0.0869. Shared source vocabulary helps the first score. The mixed labels and unused calibration split limit interpretation. vendor reportedsource

Source: Model card and weights →

A Qwen3.8-27B decision system that combines vision features with frozen PaddleOCR v5 transcriptions. OCR and image encoding run concurrently on one GPU. Published vision scores and timings cover the whole system; the internal decision backbone alone omits OCR.

Base and size

Qwen3.8-27B with D3 decision adaptations

Inputs

Text and images, including internal OCR transcriptions

Runtime

Python 3.12 and CUDA; complete-system ComboEngine or custom AutoModel loader

Licence

Apache-2.0 model tag; OCR components retain upstream terms

Access

Public weight shards, complete-system inference code and reproduction guide

• The maker reports public Decision Index Vision skill 79.7313 on 12,210 successful requests, with 263.4 ms median on H200 replicas. Loading is excluded. Private evaluation and official RTX PRO 6000 timing validation remain pending; this is not a confirmed full-board rank. vendor reportedsource

Source: Model card and weights →

A text decision model with Qwen3.5-4B backbone, adapter and a separate decision head. It supports choices, yes/no and rubric scores, with an explicit none-of-the-above option. The authors design its candidate readout for option reuse and reordered candidate sets.

Base and size

Qwen3.5-4B

Inputs

Text with supplied candidate options or scoring rubric

Licence

Apache-2.0 weights and decision head; base licence retained

Access

Public backbone, adapter and head.pt; use the Jeva project’s loading guidance

• The authors describe null-aware, permutation-invariant and cacheable candidate scoring. The checked model card has no numerical benchmark; these are design claims, not a fresh reproduction. vendor reportedsource

Source: Model card and weights →

A Laya fine-tune that labels English descriptions of simulated sensor tracks as drone, bird, aircraft or unknown. It classifies only, with no flight-control or engagement outputs. Training and testing use the same simulator with separate runs; the results say nothing about real sensors or airspace.

Base

convaiinnovations/laya, ModernBERT-large

Inputs

English text in the synthetic dataset's fixed track format

Training

10,814 rows, two epochs, soft cross-entropy; run-D selected on validation macro-F1

Licence

Apache-2.0 model card

Access

Public weights, question file and Laya loading example

• The maker reports accuracy 0.956, macro-F1 0.903 and 15-bin ECE 0.005 on 23,358 synthetic test rows from held-out simulation runs. The rule baseline scores 0.386 accuracy and abstains as unknown when evidence is thin, which the test counts as wrong for a real object. Fine-tuned unknown-class precision is 0.571. These results measure generalisation within one simulator. vendor reportedsource

Source: Model card and weights → Source: Recorded metrics → Source: Synthetic dataset →

Three sparse MobileMoE decision fine-tunes for English text or JSON. They score supplied options in one forward pass, with a pointer head and temperature scaling by question type. Weights use a noncommercial research licence; inference code is Apache-2.0. Training code and data are not released.

Size

S: 1.27B total / 0.28B active; M: 2.82B / 0.53B; L: 5.33B / 0.93B

Inputs

English text or JSON; choice, noul and score

Context

8,192 tokens per pass; split longer requests by question, reject a single question that does not fit

Licence

Weights: FAIR Noncommercial Research License and Acceptable Use Policy; inference code: Apache-2.0

Access

Public weights and MoDe inference package; custom model code requires trust_remote_code=True

• The maker reports public Decision Index 0.3 scores of 42.02, 47.10 and 51.36 for S, M and L on 140,178 requests. Four overlength requests count as wrong. These submissions have not been verified by the benchmark maintainers. vendor reportedsource
• The maker reports medians of 11.0, 17.7 and 25.0 ms for S, M and L on a half-card RTX PRO 6000 MIG partition. Each is the median of five run medians over 700 requests after ten warm-ups, using bf16 and each model’s own engine. These are not full-card leaderboard timings. vendor reportedsource

Source: Small card and weights → Source: Medium card and weights → Source: Large card and weights → Source: Inference code → Source: Licence →

An English Laya fine-tune for a Minecraft agent. Code removes impossible actions; the model chooses an intent and answers a danger question, then an executor acts and a verifier checks progress. The encoder is frozen during this training run. Its reported evaluation contains training or calibration overlap.

Base

convaiinnovations/laya, ModernBERT-large

Inputs

English game-state descriptions and valid action candidates; intent and in_danger choice questions

Training

2,564 verified decision ticks, four epochs; temperatures fitted on 284 calibration items

Licence

Apache-2.0 model card

Access

Public weights, question signatures and train_report.json; Laya runtime

• The maker reports accuracy 0.9301 before fine-tuning and 0.9325 after, with ECE moving from 0.0625 to 0.0435 on 415 items. Four items overlap training or calibration, so this is not a strictly held-out result. No score or noul items were trained. vendor reportedsource

Source: Model card and weights → Source: Training and evaluation report → Source: Agent source →

An English medical-text decision fine-tune continued from a MedGemma 1.5 4B derivative. It scores typed options without generated text. Minority-class oversampling changes the decision balance, with gains and regressions across questions. Research use only, not for diagnosis, triage, clinical decisions or patient care.

Base

unsloth/medgemma-1.5-4b-it, continued from the maker’s first released checkpoint

Inputs

English medical text and supplied typed options; text only

Training

Fresh rank-16 LoRA; one epoch, 728 steps, 51.7 minutes on an A100 80 GB, according to the card

Licence

Health AI Developer Foundations terms; derivative notice included

Access

Public merged weights, joint decision head, adapter, inference code and per-question temperatures

• The maker reports medline/adverse balanced accuracy rising from 0.625 to 0.825, above its TF-IDF baseline floor of 0.721. Macro balanced accuracy across sources falls from 0.689 to 0.686. The evaluation has 15,320 records and 19,480 decisions; balanced accuracy uses a prior-adjusted decision rule. vendor reportedsource
• The card reports PubMedQA balanced accuracy falling from 0.578 to 0.488 under that rule. Public exam sets, MeSH labels and one seed per run limit interpretation; seed variance is unmeasured. This is not clinical validation. vendor reportedsource

Source: Model card and weights → Source: Per-question results → Source: Derivative notice →

A Qwen3.5-4B LoRA for supplied medical answer choices, mainly in Chinese. It reads next-token option logits in one forward pass and returns probabilities without an explanation. This is independent of TypeSafe’s hosted Jev. The released adapter is for teaching and offline research, not clinical use or triage.

Inputs

Case text, instructions and supplied choices; mainly Chinese, with other languages represented

Limits

Released runtime: 52 options and 8,192 tokens; larger menus require the repository’s experimental tournament mode

Licence

MIT code and adapter; Apache-2.0 base, with separate upstream data terms

Access

Public custom-format LoRA and inference bundle; not PEFT-compatible

Provenance

Catch-up weights released 2 Nov; source repository published 7 Nov

• The maker reports 1,028 of 1,514 benchmark questions correct (67.90%) with the released 52-option runtime; 100 unsupported questions count as wrong. Experimental tournament handling raises this to 1,050 (69.35%). The benchmark was repeatedly used for development and is not a blind test. vendor reportedsource
• The maker reports 40 of 100 triage questions correct and uncalibrated benchmark ECE around 0.13. The original training file is lost; the public pool was reconstructed and may include sources from a later 1,557-item evaluation. That evaluation is not an untouched test of the released adapter. vendor reportedsource

Source: Model card and adapter → Source: Source and evaluation → Source: Released runtime record →

A Qwen3.5-4B derivative that reads text or images and returns probabilities for Choice, Noul and Score questions without generating an explanation. The release includes full model weights, a separate decision readout and a System One server.

Size

4.54B, including a 0.33B vision encoder

Inputs

Text or structured state and images

Image limit

1,638,400 pixels per image

Licence

Apache-2.0

Access

Public weights, Python runtime and POST /v1/systemone

Source: Card and release files → Source: Licence →

Two text decision packages. Sigma-1 merges a LoRA into Gemma 4 31B and reads option-code logits. Duo averages option probabilities from a modified KnowLine 31B and d3-ultra, both in FP8. Above 50,000 request characters, Duo uses its first member alone.

Inputs

Text state and typed choice or yes/no questions

Licence

Apache-2.0 code and weights; training sources include share-alike, copyleft-labelled and possibly non-commercial data

Access

Public weights and Decision Index engines; one 96 GB GPU in the reported setups

Limits

Sigma-1 refuses prompts above 131,072 tokens or menus above 255 options

• Logits reports a public Decision Index 0.3 score of 67.1 for Sigma-1 over 140,178 scoreable requests. Model selection used a subset of the public suite, making these scores partly in-sample. No private Full score is supplied. vendor reportedsource
• Logits reports a public Decision Index 0.3 score of 72.93 for Duo over 140,178 scoreable requests. Model selection used a subset of the public suite, making these scores partly in-sample. No private Full score is supplied. vendor reportedsource
• On one RTX PRO 6000, the maker reports median/mean request times of 122/356 ms for Sigma-1. vendor reportedsource
• On one RTX PRO 6000, the maker reports median/mean request times of 293/859 ms for Duo, with GPU memory peaking at 87,891 MiB. Its published training disclosure includes source-licence restrictions that the package licence does not settle. vendor reportedsource

Source: Sigma-1 card and weights → Source: Duo card and weights → Source: Duo source-data disclosure →

A d1-2B LoRA for checking whether a supplied answer is supported by its source. It returns one response-level support probability, without identifying unsupported spans. Use the recorded source/request/answer fields and the exact question from the card.

Size

1.9B

Inputs

English source, request and generated answer

Licence

LFM Open License v1.0

Access

Merged BF16 weights, LoRA adapter, Q8_0 GGUF and scoring records

Evaluation scope

RAGTruth; training sources capped at 12,000 characters, with longer inputs unevaluated

• The maker reports F1 81.5, AUROC 0.936 and ECE 0.030 on 2,700 RAGTruth responses. Checkpoint and threshold 0.56 were selected on document-grouped development data. The base comparison reports F1 70.9 but uses Q8_0 llama.cpp while the fine-tune uses BF16. Base-training overlap is unknown; third-party table rows use their own papers and code. vendor reportedsource

Source: Card, weights and evaluation → Source: Scoring code → Source: Licence →

A d1-2B fine-tune for choosing a replacement DOM element when a browser-test locator breaks. It reads an annotated screenshot, broken-locator details and live candidates, with an Unknown option for a missing target. The card gives no held-out accuracy or false-heal rate.

Size

1.92B

Inputs

Set-of-Marks screenshot, locator details and DOM candidates

Licence

LFM Open License v1.0

Access

Merged BF16 weights, LoRA backup and custom Transformers code

• The card proposes stopping a test for Unknown or confidence below 0.75. This is a suggested operating rule, with no published error guarantee at that cutoff. vendor reportedsource

Source: Card and weights → Source: Licence →

A CPU decision model family. Amada 3.1 uses mmBERT-base to answer yes/no, choice and ordered-score questions, with a separate probability that the text does not provide the answer for yes/no questions. Downloads pair an ONNX model with its tokenizer.

Base

mmBERT-base for 2.0 and later; ModernBERT-large for 1.0

Context

8,192 tokens for 3.x; 512 for 1.0 and 2.0

Licence

CC BY-NC-ND 4.0 for weights and repository code; non-commercial, no redistribution of modified versions

Access

Public release ZIP files and Python ONNX runner

• The maker reports 0.961 yes/no, 0.915 choice and 0.760 score accuracy on short English tests, using GPT-6 Luna labels. Human-labelled public-data results include SST-2 0.899 and Banking77 0.611. Half of the bake-off sets informed calibration, so the figures are not one untouched general benchmark. vendor reportedsource

Source: Code, model description and limits → Source: 3.1 release and weights → Source: Licence →

Three Intern-Decision-4B fine-tunes: Deem-4B for English, Karar-4B for Turkish and Wähler-4B for German. They return per-option probabilities through the System One API. Optional extensions add unknown, conformal answer sets or a short generated reasoning trace; reasoning is off by default. This is a catch-up October release.

Size

4B per language model

Inputs

Text; English, Turkish or German by model

Limits

8k context; up to 62 choice options; score uses 2 to 10 levels

Licence

Apache-2.0 code and weights; dataset terms remain separate

Access

Full weights, Q4_K_M/Q5_K_M/Q8_0 GGUF, Transformers and llama.cpp server

• The maker reports Deem-4B accuracy 94.7% and Brier 0.091 on held-out English decisions from the same pipeline as training. In 11 held-out missing-fact cases, both JevAlt and its Intern-Decision base choose unknown in 9; this does not establish a fine-tuning gain. A separate 4 Oct live test records 122/130 correct. vendor reportedsource
• The Deem Q4_K_M file is 2.71 GB; the maker measures peak RAM of 3.04 GB at 4k context. The maker reports Q4_K_M selected-option agreement with full precision of 95.2% to 97.5%, on 62 English, 61 Turkish and 40 German validation decisions. These are tested configurations, not universal hardware minima. vendor reportedsource

Source: Code, licence and serving → Source: Deem English card → Source: Karar Turkish card → Source: Wähler German card → Source: Recorded live requests →

A Qwen3.5-4B chat checkpoint with a decision readout trained through 64 added token embeddings. The original backbone weights stay frozen; a separate copy of the output table keeps the trained input rows out of chat generation. Rule questions read option-letter probabilities, while the same model file also writes dialogue. A decision still needs a prefix pass over the conversation; it is not free work inside an ordinary chat response.

Base

Qwen/Qwen3.5-4B

Inputs

Text conversations, personas and rule sheets

Runtime

Transformers; a separate GGUF package uses llama.cpp chat and completion endpoints

Licence

Apache-2.0; licence file included

Access

BF16 weights, decision protocol and GGUF variants

• The maker reports 0.958 accuracy and 0.002 ECE on 28,433 synthetic questions about rules excluded from training, with model-audited labels. A separate 1,352-question author-written set scores 0.924. The card's headline and detailed table disagree on some baseline and end-to-end figures, so no numerical end-to-end improvement is asserted here. Situation-triggered tool calls remain a stated weakness. vendor reportedsource

Source: Card and weights → Source: Licence →

A Laya fine-tune that copies a scripted lunar-lander autopilot for Low Gate. The package supplies PyTorch weights, the MLX conversion used in the benchmark, training data and hashes. It is an artifact for one simulator and prompt, with manoeuvres selected from a fixed menu.

Base

convaiinnovations/laya at revision 7b928d828b7b0e022f929d9bd2e44165aa270148

Inputs

Text game states and manoeuvre questions

Licence

Apache-2.0; base licence and modification notice included

Access

PyTorch and MLX checkpoints, data and training report

• The maker reports 142 landings out of 160 levels on courses 5001-5020, excluded from training and tuning. The training report separately contains 2 overlapping items among 2,728 evaluation rows. PyTorch and MLX choose the same manoeuvre on 5,795 of 5,801 recorded states; this open-loop parity check does not fly new courses. vendor reportedsource

Source: Card and checkpoints → Source: Training report → Source: Low Gate benchmark →

A Qwen3.5-4B LoRA fine-tune with a pointer head that scores markers beside the supplied candidates. Multiple questions about one state share a sequence. The pointer readout removes the letter-based 26-option cap, but all input still has to fit the runtime. The model selects among supplied options and has no abstention mechanism.

Base

Qwen/Qwen3.5-4B; merged BF16 LoRA

Inputs

Text state and candidate questions; Chinese and English

Readout

256-dimensional pointer head; fitted temperature 1.538

Licence

Apache-2.0 weights; source datasets retain their own terms, including non-commercial or research-only terms

Access

Merged model.safetensors with loading and evaluation code

• The maker reports public Decision Index 0.3 skill 61.69 and raw 71.01 over 140,178 requests. The saved scores contain 140,177 ok responses and one unsupported response. Candidate-order debiasing averages original and reversed orders, so the reported 64.2 ms median and 147.3 ms mean on an A800 include that recipe. The maker describes 13-word overlap filtering; this is not an independent contamination audit. vendor reportedsource

Source: Card and weights → Source: Scores and request counts → Source: Serving and training code →

A full-parameter Qwen3.8-27B vision decision fine-tune with a separate head and a frozen evaluation engine. A forward pass returns probabilities across supplied choices without generating text. The released package includes weights, inference code, manifests and full public vision-suite records.

Base

Qwen/Qwen3.8-27B

Inputs

Text and images; 2 to 26 choice options

Limits

262,144 input tokens; up to 1,638,400 pixels per image, subject to device memory

Licence

Fine-tuned weights CC BY-NC 4.0; original backbone Apache-2.0

Access

BF16 backbone, FP32 stored head and single-A800 inference recipe

• The maker reports public Vision Index 76.7057 with 9,867 of 12,210 requests correct. Pooled across eight parallel processes, each on one A800 80 GB, synchronized forward timing has 361.5 ms median, 507.2 ms mean and 581.9 ms p80, excluding preprocessing and transfers. The official private evaluation and RTX PRO 6000 timing check are pending; this is a self-scored public result. vendor reportedsource

Source: Card, code and weights → Source: Pinned public results → Source: Weight licence →

A Qwen3.5-9B LoRA decision model with a separate 255-option head. The pinned V1.6 package returns one zero-based choice and ordered probabilities through POST /decision. It needs the public custom loader; it does not generate explanations or tool calls. The package predates this scan window, while the availability clarification appeared during it.

Base

Post-trained Qwen/Qwen3.5-9B; rank-64 LoRA and a separate BF16 decision head

Inputs

Text context and 2 to 255 ordered, non-empty options

Limits

8,192 rendered tokens; no silent truncation

Licence

Apache-2.0 project code and derivative weights; upstream components and datasets retain their terms

Access

Pinned complete V1.6 package at revision 545549b8e8104f1780966ef1c13136abb32712b4; custom Python 3.13 loader

• The maker reports public Decision Index 0.3 skill 49.83 and balanced raw score 62.28: 140,299 of 140,620 requests answered, with 321 unsupported and no errors. Development data selected update 71,327, so its 254/352 result is not an independent final test. V1.6 uses temperature 1 without fitted calibration. No private Full score or official rank is asserted. vendor reportedsource
• The card reports operator-run tensor parity and three cold CUDA research/public-loader checks with zero maximum probability difference. This scan checks the pinned file inventory, manifest hashes and checkpoint metadata; it does not repeat GPU inference or verify an independently published parity receipt. vendor reportedsource

Source: Current V1.6 card → Source: Pinned package → Source: Loader and licence → Source: Package manifest →

A 3.93B full fine-tune of Domyn-Edge-SFT for text decisions. It reads probabilities over option-letter logits in one forward pass. The card provides Transformers and vLLM recipes; the served vLLM path requests one output token. Its public JevBench submission is awaiting a sealed evaluation.

Base

domyn/Domyn-Edge-SFT; 3,928,787,968 stored parameters

Inputs

Text state and Choice, Noul or Score questions

Context

16,384 tokens used in training, per the card

Access

FP32 safetensors and custom Transformers code

Licence

Apache-2.0 declared in the card; no separate package licence file

• The current card reports public-only composite 78.78 and claims second place in JevBench v1.8.7, with sealed evaluation pending. These are the maker's tables; no official board placement is established here. The bundled eval_report.json remains a separate 6,355-item report with raw accuracy 0.8017. vendor reportedsource

Source: Card and weights → Source: Evaluation report → Source: Pending JevBench adapter →

A fine-tune of Liquid AI’s d1-omni-600M for a Clash Royale course project. It reads a text game state and up to four candidate plays, including tactical notes, retrieval and knowledge-graph context. The package selects a play; the surrounding agent uses iPhone Mirroring on a Mac. Intended for research and the course demo in Training Camp or friendly battles.

Base

LiquidAI/d1-omni-600M; 587,161,089 stored parameters

Inputs

Text game descriptions and up to 4 candidate plays

Training

3,641 four-option questions from 197 recorded professional games; splits by video

Access

FP16 weights and bundled custom loader; CPU loading uses FP32

Licence

LFM Open License v1.0; file included

• The maker reports four-way accuracy 0.507 on 766 test lines from 27 held-out games, against chance 0.25 and a placement prior 0.384. The reported 95% interval is 0.473 to 0.540. FP16 and FP32 choose the same option on 100 check lines. These are offline play-choice results, not a measured live-game win rate. vendor reportedsource

Source: Card, weights and code → Source: Licence →

Two pruned QEV v3 models for text and image questions. The maker removes eight of 24 language-model blocks and trims the input vocabulary to 32,768 rows, then distils a decision readout. They return option probabilities, a typed answer and a separate abstain flag. Ken Jo's QEV family is separate from labqoat's Qev 0.8B.

Sizes

466,313,408 and 1,314,213,824 parameters, per the cards

Inputs

English text and up to four images; Choice, Noul and Score

Context

8,192 input tokens; over-budget requests refused

Runtime

Bundled backbone and head; pinned qev Python loader on CUDA, accurate mode only

Licence

Apache-2.0 files included; training datasets retain their own terms

• The maker reports public vision index scores of 52.52 and 62.15 on 12,210 rows across 11 benchmarks, versus 52.70 and 62.27 for the corresponding full-depth parents. These are public-only normalized scores, not accuracy percentages; private-half results are pending. vendor reportedsource
• On an RTX 4060 Ti 8 GB at batch 1, the maker reports peak GPU memory of 1.15 GiB and 2.76 GiB, with median p50 timings of 54.4 ms and 99.0 ms across 47 request profiles. Calibration and abstention thresholds were fitted on development data; the empirical error target is not a guarantee on new data. vendor reportedsource

Source: 2B compact card and weights → Source: 0.8B compact card and weights → Source: Licence →

A 144.3M-parameter multilingual encoder adapted from mmBERT-small to score supplied options. It accepts named Choice, Noul and ordered Score questions, with a Python runtime for CPU or CUDA. This September release was found through an overnight card edit.

Size

144,292,870 stored parameters; FP32 weights

Inputs

Text state and 2 to 20 options per native question

Context

8,192 combined tokens supported by the runtime; historical accuracy tests used 1,024

Access

Open weights and packaged Python inference code; private training pipeline

Licence

Apache-2.0 declared in the card; no separate package licence file

• The maker reports 1,463/2,000 typed questions correct on H200 BF16. A separate CPU FP32 reproduction gets 426/600 Choice, 542/800 Score and 483/600 Noul. Jev figures in the card are supplied references, not newly matched calls; an 8,192-token smoke test does not establish accuracy at that length. vendor reportedsource

Source: Card, runtime and weights → Source: CPU results → Source: Checkpoint provenance →

A LoRA adapter and option-scoring head on Qwen3.5-2B-Base for Choice, Noul and Score questions. This v1 release averages three training runs and supplies a System One server. Its optional vision path keeps the base vision tower without image training. The community SageMaker project deploys the separate Strands Decider 2B hobson-v19 checkpoint.

Base

Qwen/Qwen3.5-2B-Base; separate base download required

Inputs

Text; optional image path needs the vision extra and a runtime release after 0.1.0

Runtime

strands-decider Python package, CPU, MPS or CUDA; POST /v1/systemone

Licence

Apache-2.0; LICENSE.md included

• The maker reports 180/231 public JevBench tasks correct at a 4,096-token serving window, Brier 0.280 and ECE 0.072. CPU and NVIDIA L4 select the same answers on all 231 tasks; the training harness differs on one near tie. These are maker runs, not an independently reproduced board score. vendor reportedsource

Source: Card, adapter and run records → Source: Licence → Source: Runtime and training code →

A CPU decision model that scores supplied options against a fixed zero-energy none state. It packages an encoder derived from multilingual-e5-small, ONNX weights and a fitted head. Its runtime sorts options into a canonical order before scoring, so order consistency belongs to the complete serving path.

Size

89 MB model files, per the maker

Inputs

Text and option labels; no free-form generation

Limits

Reads 192 text word pieces and 16 per option, with truncation flags; options processed in blocks of 20

Access

Bundled model files; Python with NumPy, ONNX Runtime and tokenizers

Licence

Apache-2.0 code; upstream multilingual-e5-small is MIT, with notices

• The maker reports 53.6% correct decisions including none across 900 matched cases, half with no right option, versus 58.6% for Strands Decider 2B v1. ZeroPoint takes a 42 ms median on a two-core CPU with two threads; Strands uses an NVIDIA L4 in that comparison. These timings do not compare raw speed on the same hardware. vendor reportedsource

Source: Code and bundled model → Source: Saved comparison → Source: Licence → Source: Model manifest → Source: Package releases →

Medical decision models in two sizes, built on Qwen3.5-4B and Qwen3.8-27B with a trained decision head. They read text and images and return probabilities for each supplied option. These are research models; the card requires independent validation and clinician review before patient decisions.

Sizes

4B and 27B backbones; the 4B package stores 5,287,287,300 parameters, including a copy of the tied output embedding

Inputs

Text or JSON with optional images; up to 64 questions and 4,096 prompt tokens

Access

BF16 weights, custom Transformers code and a Python request wrapper

Licence

Apache-2.0 files included; source datasets retain their own terms

Memory

The maker reports about 16 GB GPU memory for 4B and 60 GB for 27B

• The maker reports 86.1% accuracy for both sizes on the same 6,700-question medical comparison subset. On its 554 out-of-domain questions, 4B scores 71.5% and 27B 74.4%. These are constructed benchmark questions, not prospective patient outcomes. vendor reportedsource
• The maker reports median request latency of 68 ms for 4B and 148 ms for 27B on one H100 at batch size 1. Hosted-model timings in the card exclude client networking, so they are not end-to-end comparisons. vendor reportedsource

Source: 4B card and weights → Source: 27B card and weights → Source: 4B configuration → Source: Licence →

A small Russian/English decision model with a custom hybrid backbone and decision head. The 1110 checkpoint adds toxicity and profanity training. It scores supplied options, with reversed-order averaging enabled by default; that serving choice uses two orders.

Size

The card lists 29.7M parameters plus 16.8M n-gram parameters, despite the 26M name

Inputs

Russian and English text; choice, yes/no and score questions

Limits

Reads the first 1,536 state tokens; training-family wording is recommended

Access

Transformers custom code; safetensors and QPACK files

Licence

Q Project Model License 1.0: free non-commercial use; commercial use, including internal use, needs a paid licence. Company evaluation is allowed for up to 90 days without production or customer use.

Data terms

The card identifies non-commercial ANLI data and toxicity data without a stated licence; third-party terms still apply

• The maker reports 45.9 on the 231-task public JevBench v1.2 run, down from 46.3 for 1027. Its reported Russian/English toxicity accuracies are 96.1%/90.1%; harmless English uses of violent words remain failure cases. These task results do not establish general reasoning ability. vendor reportedsource

Source: Card and weights → Source: Licence → Source: Runtime →

An English game-scene decision model derived from Ettin-400M, with ONNX files and a Python runtime. It scores NPC intent, mood and quest questions, with optional abstention and 42 compiled adapters. The runtime adds option-order averaging and a minor-safety guard; raw ONNX does not include either.

Inputs

Text or JSON game scenes; choice with 2 to 32 options, score with 2 to 10 levels, and yes/no

Limits

3,072 state tokens; two-option choices and choices with 11 or more options never abstain under the published profiles

Package

792.5 MB default fp16-weight ONNX file; Python runtime 0.1.0

Licence

Apache-2.0; upstream Ettin-400M is MIT

Scope

Fictional game situations; the card excludes consequential decisions about real people

• The maker reports 60.0% accuracy on 2,993 private questions from unseen task families with option-order averaging. Default thresholds cover 66.2% of that set, versus 97.4% of dev questions. The private evaluation sets and training rows are not released. vendor reportedsource
• On an AMD Ryzen 5 3600X, the maker reports warm model p50 0.92 s for one question in a single option order and 1.80 s with default choice-order averaging. The 200 held-out rows have a median 325 tokens. vendor reportedsource

Source: Card and files → Source: Training run record → Source: Data sources → Source: Licence →

A text decision package with the MNN-Pocket-0.3B backbone, an unmerged LoRA adapter and a pointer head. It returns probabilities for supplied choices, yes/no questions and ordered scores. The bundled Transformers/PEFT runtime keeps the full backbone and language-model head.

Package

723 MB backbone, 28.7 MB adapter and 1.8 MB decision head, per the card

Inputs

Chinese and English text; up to 255 candidates per question

Context

32,768 request tokens; longer requests raise an error instead of truncating

Access

Bundled Python runtime; CPU float32 or supported GPU dtypes

Licence

Apache-2.0 file included

• The maker fitted a temperature of 1.1112 on a 600-question calibration partition, reducing pooled NLL from 0.8598 to 0.8576. This calibration fit does not establish held-out task accuracy or phone performance. vendor reportedsource

Source: Card and weights → Source: Runtime → Source: Calibration record → Source: Licence →

A Qwen3.5-0.8B-derived router that predicts model success and token use separately from the prices supplied by an application. The package includes routing and content-signal heads plus a general question head. Sage and Swift are policies on the same model.

Base

Qwen/Qwen3.5-0.8B; specialized heads use the lower 12 layers, while the question head uses all 24

Inputs

Agent conversation and developer-defined questions with answer options

Access

Open weights and hosted arc-sage-1.0 / arc-swift-1.0; public vllm-arc plugin is unfinished

Serving

Stock vLLM cannot load the architecture alone. The public plugin supports encoder integration and Stage-1 scoring, without the complete policy, question or content-signal APIs

Licence

Apache-2.0; LICENSE and NOTICE included

• The maker defines utility as value of success times predicted success minus expected cost. Sage uses value 4.5 and Swift 0.1. Prices are supplied at runtime; the package contains no price snapshot. vendor reportedsource

Source: Card and release files → Source: Policies → Source: Provenance → Source: Licence → Source: Public plugin and limits →

A four-size family with Qwen3.5 hybrid backbones and native decision heads. It scores supplied candidate meanings for Choice, Noul and Score questions without generating text. Published decision results are text-only; retained vision components do not establish image accuracy.

Sizes

0.8B, 2B, 4B and 9B releases; the 2B card lists 2,319,227,200 stored parameters

Runtime

Custom Transformers code with AutoModel and trust_remote_code, or the GS1 HTTP server

Context

Configured branch limits: 65,536 tokens for 0.8B/2B/4B and 32,768 for 9B; default HTTP service limit is 4,096

Calibration

Released configurations use identity calibration; task-specific confidence still needs validation

Licence

Apache-2.0 code and weights; LICENSE and NOTICE included

• The maker reports GS1-2B scoring 62.22 on public Decision Index 0.3 across 37 benchmarks, and 80.09% on 231 public JevBench decisions. Public task accuracy is separate from the sealed composite. vendor reportedsource
• The maker reports GS1-2B warm in-process CUDA Graph p50 18.64 ms on an RTX 4090 reporting 48 GiB, using one question and a 256-token longest branch. Loading, graph capture, HTTP and SDK transport are excluded. vendor reportedsource

Source: 2B card and code → Source: 0.8B package → Source: 4B package → Source: 9B package → Source: 9B request limits → Source: Licence →

Trained candidate-scoring heads for frozen Qwen3.5-0.8B and 4B backbones. Each package contains the head, temperature calibration and an execute-or-escalate policy, with a separate backbone download. The English/Italian training and evaluation situations are synthetic.

Package

Head-only releases; ANT-08B has 1,053,696 trainable parameters and a 4.2 MB head, excluding the backbone

Inputs

Text state, question and supplied candidates; configured context budget 1,024 tokens and 128 per candidate

Runtime

Python runtime from the matching ant-08b or ant-4b GitHub repository; optional MLX path

Status

Work in progress; one training run per size, without measured seed variance

Licence

Apache-2.0 for released heads; Qwen backbones retain their own Apache-2.0 licence

• The maker reports test accuracy 67.44% for ANT-08B and 73.92% for ANT-4B on 648 synthetic situations. On 456 situations from held-out use cases, accuracy is 57.02% and 59.65%. These are separate from the validation-fitted escalation policy. vendor reportedsource
• ANT-08B reports validation coverage 44.2% at 9.9% error among accepted decisions, using 688 validation rows. Its fitted threshold does not establish that error rate on new domains; OOD ECE is 0.220. vendor reportedsource

Source: ANT-4B card → Source: ANT-08B card → Source: 4B metrics → Source: 08B policy → Source: Licence →

A Japanese decision adapter and pointer head trained on Sarashina2.2-0.5B with the Strands Decider recipe. It returns Choice, Noul and Score answers through strands-decider. This is a community release, with a separate base download or a merged Q8_0 GGUF.

Base

sbintuitions/sarashina2.2-0.5b

Inputs

Japanese text; training and request configuration use a 4,096-token maximum

Calibration

Temperature 1.0, without calibration; the card warns about long documents, arithmetic, neutral labels and surface-word effects

Training

200,156 rows for one epoch; translated public data and synthetic data. The mixed training rows and teacher distributions are not published

Licence

MIT; LICENSE.md includes the base and Strands code attribution

• The maker reports median inference 0.94 s for 224 tokens and three questions on AWS Lambda arm64 with 10,240 MB memory, six Graviton2 vCPUs and Q8_0 GGUF. A separate roughly 220-token one-question test reports median 0.38 s and p95 0.58 s. vendor reportedsource

Source: Card and packages → Source: Training configuration → Source: Request configuration → Source: Licence →

A text-only Gemma 4 31B fine-tune that scores supplied option tokens without generating an answer. It reads complete inputs twice when they are at most 3,000 tokens before repetition, handling each question separately. The package includes full BF16 weights and its inference code.

Base

google/gemma-4-31B-it; full-parameter backbone training with token embeddings and output head frozen

Inputs

Text Choice with 2 to 255 options, or Noul; up to 64 questions per request

Context

262,144 rendered tokens per question, including repetition and template; unsupported questions reject the whole request without truncation

Access

16 BF16 weight shards, approximately 62.55 GB before runtime overhead; CUDA package and CLI

Licence

Apache-2.0 release; retain base-model and MIT helper notices

• The maker describes training on more than 2 million cleaned text-decision examples with teacher-generated soft labels. The public score is reused from an earlier full run; publication checks do not constitute a fresh full benchmark. vendor reportedsource

Source: Card and inference package → Source: Licence → Source: Third-party terms → Source: Weight manifest →

A Chinese/English post-trained Qwen3.8-27B decision model with full BF16 weights. Its decide method scores a 26-slot decision head once per question and returns Choice, Noul or Score answers, without generating text. Its reported N-bank evaluation was also used during interface development.

Base

Qwen/Qwen3.8-27B; upstream revision recorded in provenance.json

Inputs

Text or JSON state; 1 to 64 questions; Choice has 2 to 26 options and Score has 2 to 10 ordered levels

Context

4,096 tokens per rendered question; longer inputs raise an error rather than truncate

Access

18 BF16 weight shards, about 52.17 GB before runtime overhead; custom Transformers code with trust_remote_code

Calibration

Probabilities are conditional on supplied alternatives; downstream-domain calibration is unmeasured

Licence

Apache-2.0; actual LICENSE included

• The maker reports 1,311/1,340 classification and Boolean decisions correct on the development N bank. Its 120 tied mixed decisions are excluded from accuracy; the 300 rating prompts use a separate MAE metric. This is a same-bank regression check. vendor reportedsource

Source: Card and weights → Source: Licence → Source: Provenance → Source: Evaluation record → Source: Typed inference implementation →

A provisional Italian sentence classifier fine-tuned from Laya multilingual. It finds explicit statements of missing knowledge or experience in interview-practice answers. This is a fixed two-class encoder and head, separate from Laya’s general typed-question heads; it does not assess competence or choose interview moves.

Size

306,941,186 parameters; full float32 weights and a 308 MB dynamic-int8 ONNX export

Inputs

One answer sentence paired with the last 300 characters of the interviewer’s question; 192-token truncation

Runtime

ONNX Runtime, Tokenizers and NumPy on CPU; no Laya SDK required

Training

2,549 synthetic Italian sentence pairs, 362 positive; final checkpoint trained on all rows

Calibration

Threshold 0.5; no calibrated confidence guarantee or real-interview acceptance test

Licence

Apache-2.0 fine-tuning and exports; retain Laya and mmBERT upstream attribution

• The maker reports 288/294 positive answers found and 8/312 negative answers flagged in a synthetic leave-one-profession-out pilot. Those results concern held-out-group models, not an unseen test of the final checkpoint trained on all rows. vendor reportedsource

Source: Card and weights → Source: Licence → Source: Checksums and provenance → Source: Pilot and integration record →

A multimodal decision API for text, images, audio, video and PDFs. It returns yes/no probabilities, a choice or a rubric score without generating text. Questions share the input but cannot see each other’s answers; application code combines the results.

Price

Google reports $0.10 per million input tokens; output free. Shared input billed once, plus each question's instructions and criteria · Compare prices

Inputs

Google reports support for text, images, audio, video and PDFs; the 262,144-token context includes shared input and all questions

Questions

Google reports up to 1,000 independent questions per request; 2 to 255 choice options or ordered score levels

Access

Gemini API in Google AI Studio, including a no-cost prototyping tier; listed on OpenRouter

Limits

No text generation, function execution, caching or Batch API support. Google reports yes/no probabilities from 0.01 to 0.99

Media billing

Google’s launch lists no separate audio transcription or OCR charges

• Google reports 150 to 250 ms end-to-end latency for inputs up to 16K tokens, nearly flat from 1 to 100 questions. Those measurements do not cover the maximum context or question count. vendor reportedsource
• Singularity Feed’s independent test asks 1,000 yes/no questions over one 1,000-sentence input. The author reports 92.5% accuracy and median 3.996 seconds across five requests, with estimated cost $0.006743 per request. This tests a shared-input workload, not 1,000 separate documents. independentsource

Source: Launch and billing → Source: API guide → Source: Model limits → Source: OpenRouter listing → Source: Independent test, code and per-answer records →

A small typed-decision encoder for text, with a separate femto-asr speech frontend. It scores options in a forward pass per question. The maker’s speech tests show that audio frames help on familiar recording conditions but add no benefit on unseen phone speech.

Size

Maker reports 7.7M decision-model parameters and 30 MB fp32 weights; speech frontend adds 2.39M parameters

Inputs

The maker reports text or 16 kHz mono speech input through femto-asr; choice, yes/no and score questions

Access

Public weights and single-file Python inference; no text generation

Licence

Apache-2.0, with upstream notices

Limits

English; out-of-list spoken commands are not reliably rejected. Confidence is not an out-of-distribution rejection guarantee

• The maker reports 73.2% accuracy across 400 typed-decisions test cases containing 2,000 decisions. The model trained on that dataset’s training split. SLURP accuracy is 71.4% with transcript plus frames, against 58.6% with transcript alone; on unseen MInDS-14 phone speech the figures are 55.8% and 56.1%. vendor reportedsource

Source: Card and weights → Source: Inference code → Source: Licence → Source: Configuration →

A Laya fine-tune that assigns clinical notes to a fixed menu of medical-specialty and document-type labels. Snapcall supplies the loader and System One server. Its published test concerns document classification; it does not validate diagnosis or clinical care.

Base

convaiinnovations/laya, fully fine-tuned with Snapcall

Inputs

AI Anytime reports text-only choice classification over 40 supplied labels

Access

Public backbone and decision-head safetensors; load or serve with Snapcall

Training

AI Anytime reports 4,023 training, 447 validation and 496 test examples from galileo-ai/medical_transcription_40

Licence

Apache-2.0 according to the card; dataset terms remain separate

• AI Anytime reports 64.31% test accuracy, macro-F1 0.5706 and 15-bin ECE 0.0480 on 496 examples. Temperature scaling uses the validation split; calibration must be checked again on new traffic. vendor reportedsource

Source: Card and weights → Source: Evaluation record → Source: Training and serving toolkit →