Laya is Convai Innovations' decision engine on ModernBERT-large with 421M parameters. A 322M multilingual sibling is also available.
By Convai Innovations · Released 18 Sep 2026
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.
Laya: A non-autoregressive decision engine on ModernBERT-large (421M), with a 322M multilingual sibling.
Laya: Convai Innovations
The 9 November 2026 card adds a citation to From Laya to Vega: Open Non-Autoregressive Decision Models with Calibrated Probabilities. The recorded change edits the card; it does not upload new checkpoint weights. The paper’s results have not been audited here.
Source: Site → Source: GitHub → Source: Model card and preprint citation →
A local inspection workbench: OpenCV measures the image, hard rules decide anything numeric, and a Laya decision model cross-checks the result. Built for fruit picking, weed and pest monitoring, and steel pipe and textile checks. Version 3.0 also trains your own Laya-format decision models.
A native DuckDB extension that adds seven Jev-style SQL functions backed by a local Laya model: DuckDB handles geometry and distance, Laya reads the text fields. It uses no hosted model or API keys.
Runs Bespoke Nimble-9B and three Laya checkpoints as players over the same 200 seeds of a headless Slay the Spire simulator, zero-shot and fine-tuned. The results are the authors' own.
Classifies Apple Mail messages as for you, to review or junk with a Laya model running on your own computer, and shows a report of what it would do. Mail never goes to an external AI service, and it never deletes anything.
Routes support messages through regex, then a small classifier, then decision models, then an LLM as a last resort, all running locally. Laya answers the guardrail and department questions, and Tev1 is the default for the next tier.
Tags each pet product's species, life stage and allergens from the store's own text, asking Jev or Laya one typed question per value: supported, contradicted, unknown or conflicting, with the source sentence attached. Merchants can approve, correct and lock tags, or let a validated policy accept them automatically. A hackathon entry for AIBoomi Startup Weekend Kochi, 23-24 Oct.
An MCP server that gives an AI assistant yes/no, choice and score tools backed by whichever decision model you configure: any /v1/systemone endpoint such as Jev or Clef, or Laya running on Apple silicon. Its tools never name a model family.
A small ticket router: with the Laya backend selected, each support ticket goes to Laya, the open Jev-style model, with three questions (which team, how urgent, is it spam). The default backend is a keyword stub, with a slot left for Jev.
Simulates a moderation queue hit by a wave of privacy-invading posts the existing router did not prioritise: Claude reads 20 of those posts and 20 ordinary ones and drafts one yes/no question, and a fine-tuned version of the open decision model Laya asks it of every new post so the matches jump to the urgent queue.
jevgrep finds code by asking Jev thousands of yes/no questions. layagrep records those questions on SWE-bench tasks and fine-tunes the open Laya to answer them locally, with no API key and no code leaving the machine. A smoke run on 400 held-out questions lifted file-navigation AUC from 0.79 to 0.94 (Jev 0.97), but declaration relevance stayed at 0.58 (Jev 1.00).
A desktop notes app for Mac, Windows and Linux that keeps plain Markdown files and runs the open Laya model on your own computer. Each tag is a yes/no question, like "Is this note about a trip?"; Laya answers all of them for a note in one pass and keeps the tags it is confident about. Notes stay on the computer; the only network request is a one-time download of about 1.6 GB of model files.
Runs multilingual Laya through a local Go/ONNX service, with a ZCode plugin for risk checks, context selection and 13 MCP tools. VENDOR-REPORTED on an RTX 3060, three runs per variant: total input fell from 359,085 to 239,544 tokens (33.3%) and requests from 16 to 9, but non-cached input rose 21.8%. Laya ranking alone cut total input 9.3% and non-cached input about 2.8%, within the observed range; the author has not established a fresh-token or cost saving. The /decide endpoint took about 26 ms at p50.
Checks coding-agent tool calls with rules and a local multilingual Laya fine-tune, returning allow, review or block. VENDOR-REPORTED, base to fine-tune: the model alone held 32 then 39 of 43 attacks and allowed 25 then 44 of 50 everyday commands. Six end-to-end verdicts improved and two worsened, missing a preset no-regression criterion. Accuracy on 162 held-out training rows rose from 0.611 to 0.852. The author inspected the benchmark and evaluation sets between training runs, so they are less independent than a fresh test.
A local laya-multilingual fine-tune labels 9,696 Bilibili comments about Phantom Blade Zero with 14 questions and a two-level topic tree. The author reports 60.8 minutes of training on an RTX 3060. On 771 test comments with 1,485 decisions, topic accuracy rises from 0.210 to 0.911 and validity from 0.253 to 0.971. Test labels come from rules, whose topic precision is estimated at 72% to 75%, so 0.911 measures agreement with the rules. The README also discloses some train/test overlap. Toxicity detection fails at macro-F1 0.489, with only 11 training positives.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Scans three Laya checkpoints for answers that ignore the state, option-order changes and negation failures. In the author's 04 Nov scan, 5 of 984 instructions collapse on English Laya, 5 on typed-decisions and 35 on multilingual. On a 216-answer realistic guardrail panel, 16, 6 and 42 answers respectively are wrong at confidence of at least 0.9. The panel uses six texts per word, and some scan questions are semantically odd. Results identify cases to check, rather than an overall model ranking.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Runs Laya locally in a browser through ONNX Runtime Web, using WASM on CPU by default. The author reports 97.9% top-answer agreement with PyTorch across 48 questions for each quantized build. In Chromium, the default q8e8 build matches 11 of 12 top answers. A separate synthetic 108-case test of the original PyTorch model gets 59.3% of 312 answers right; that measures model behaviour, not browser parity. Quantized confidence values are approximate because calibration used the full-precision model.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Source: Synthetic test and limits → Source: Apache-2.0 licence →
Stores versioned decisions in Postgres and checks calibration against the user's labels. The author's Phase 0 run used Jev 1.13 on 2,000 consumer complaints for about $0.30. Top answers changed on 4.8% of rows across three identical runs, and on about 16.5% when the same rows were rebatched. The repository includes measured Phase 1 ledger and routing runs but has no published release. Its claims about managed hosts have not been tested on Supabase, Neon or RDS themselves.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Decision models choose one of 35 tilt-and-throttle manoeuvres on the same seeded lunar-lander courses. The game waits for the answer, so measured latency never changes the flight score. Across 160 levels per lane, the author reports 159 landings for a scripted autopilot, 156 for Jev, 151 for Kev-27B and 142 for the Laya fine-tune. Prompt selection and Laya training are documented; these results cover this simulator, with recordings available for replay.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Source: Recorded benchmark → Source: Prompts and training → Source: Licence →
Tests whether LLM-written weight edits or Python overrides repair Laya's routing of Banking77 messages into five departments. In a two-seed pilot, the best code-patch mean is 78.1% accuracy, against 76.9% for the best weight-edit mean at a different effort setting and 70.3% unedited. Test-selected fine-tuning reaches 75.0%. Budgets differ beyond label counts, and isolation relies on paths and log scans. Stored logs and summaries support the reported runs; saved edited weights are omitted.
Authors' published runs · VENDOR-REPORTED; not independently reproduced
Rules handle unsubscribe and automatic replies first; local Laya labels the rest, with uncertain messages sent to an LLM. The author reports 110 of 117 selected replies correct at a 0.80 threshold out of 182 real replies. Rules were written after reading that same set, and model-drafted labels remain under review. Fresh held-out performance is unknown.
Published author-run records, checked 5 Nov 2026; not reproduced in this scan.
Source: Licence file → Source: Integration code → Source: Published run →
A Korean adaptation of layaForWorkflows, using the multilingual Laya checkpoint in the browser. It records model answers for 37 examples across seven workflows. At branch threshold 0.50, its replay routes 15 automatically, 17 to people and five to blocking. Wording was adjusted against these examples, so this is a demonstration set.
Published author-run records, checked 5 Nov 2026; not reproduced in this scan.
Source: Licence file → Source: Integration code → Source: Published run →
Runs multilingual Laya and a multilingual MiniLM embedder in the browser on Korean fictional documents. Its Korean report gives 62.5% accuracy on 72 claims with the default two-option question and 28% final-outcome agreement on 32 routing questions. The retained English README describes upstream results, not these Korean runs.
Published author-run records, checked 5 Nov 2026; not reproduced in this scan.
Source: Licence file → Source: Integration code → Source: Published run →
Labels spreadsheet text locally with Laya and exports answers, confidence values and review flags. On 300 texts per dataset, the author reports 93% English news-topic accuracy, 68% Portuguese review sentiment accuracy and 33% exact star-rating accuracy. The review labels come from customer ratings, which can disagree with the text.
Published author-run records, checked 5 Nov 2026; not reproduced in this scan.
Source: Licence file → Source: Integration code → Source: Published run →
A Korean browser adaptation of Laya Memory, using multilingual Laya and MiniLM to flag stale stored facts. On 62 hand-written cases, the author records 19.4% strict accuracy and 27.4% lenient accuracy with the default policy; it closes one of 22 stale facts and wrongly invalidates one standing fact. This sample also shaped the questions and thresholds. App-event pairs have not been recorded.
Published author-run records, checked 6 Nov 2026; not reproduced in this scan.
Source: Licence file → Source: Integration code → Source: Published Korean run → Source: Recorded answers →
A Telegram bot that checks spam and links related messages before deleting them, with admin exemptions and a keep-on-failure policy. Its published evaluation records 75 live Jev calls, including two rejected prompt changes. Those calls did not exercise Telegram deletion. The README says its production classifier is now a Laya fine-tune distilled from Jev.
Licence, implementation and recorded live model calls checked 6 Nov 2026; no fresh Telegram run.
Source: Licence file → Source: Integration code → Source: Recorded model calls →
A browser element picker tested on an M2 Pro with 16 GB RAM. The author records 94 correct element picks from 94 live-site decisions across five runs, while only 91 of 95 browser steps finish. Page and action failures explain the difference. The wider grounding comparison finds no significant accuracy difference after multiple-comparison correction. Raw runs and a demo are public.
Author-run records and source checked 7 Nov 2026; not rerun locally.
Source: Licence file → Source: Grounding and browser code → Source: Recorded live run → Source: Results and limits →
Runs Laya through native MLX, WebGPU or a TypeScript CPU reference. The author records 756 of 756 matching top answers against Python across three checkpoints and four GPU configurations. This measures implementation agreement on 63 questions, without labelled task accuracy. Q4 changes one to five top answers per checkpoint on those questions.
Author-run parity and performance records on a fanless Apple M2. Matching top answers does not mean identical probabilities; sustained thermal throttling changes timing.
Source: MIT licence → Source: Agent code → Source: Parity and timing records → Source: Recorded Python outputs →
Moves llama.cpp decision serving code into a C library. Its recorded d1 request takes 678 ms on eight CPU threads. Saved CLI/server parity tests stay within 0.01 for Laya, exceed that probability tolerance for H2O and Clef, and reject the tested Perplexity package. On a Pixel 8a in Termux, the maker reports a 464 ms yes/no decision with d1-2B Q4_K_M on four threads and 2.1 GB RSS. Without repacking, that question takes 1,210 ms at 1.1 GB RSS. iOS and Capacitor app bindings remain plans.
Source: Licence → Source: Saved parity output → Source: Phone measurement and build notes →
Local CPU inference through laya-ts and ONNX, exposed as MCP tools for typed text questions. The maker records a thank-you note changing urgency when options are reordered. Optional stability checks re-ask the question and flag flips; their scores are uncalibrated, and stable answers can still be wrong.
Author-run records; not independently reproduced
Source: Licence → Source: Model integration → Source: Real-model test path →
Selects transcript windows with text-only Laya-multilingual and deterministic features, then asks Ollama for titles and metadata. The maker records real Hindi/English Laya calls and a synthetic end-to-end render. Hook questions disagree on one clip. Live target-machine Whisper GPU and Ollama stages remain untested in that record.
Author-run records; not independently reproduced
Source: Licence → Source: Recorded execution and limits → Source: Laya integration →
A CLI and dashboard for preparing labelled examples, training decision heads, evaluating predictions and serving a System One endpoint. Its released Laya medical-document classifier includes a separate held-out test report. The toolkit’s RLCD recipe is its own implementation; it does not disclose TypeSafe’s training code.
Public trained weights and evaluation record; no fresh training run
Source: Licence → Source: Released model and report → Source: Held-out evaluation →