Laya

Last updated: October 20, 2026

Open30–80ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

Maturity6.2/10
Apache weights plus PyPI / HF / Router packaging and clear docs lead open maturity; since 2 Oct also served by llama.cpp's /v1/systemone (ggml-org GGUF). Self-host only (no hosted SLA) caps the top.
Capability5.5/10
Official boards (17 Oct): Decision Index 0.3.2 Full 4.43 (#164); JevBench Intelligence 1.8, calibration 63.2, p50 0.825 s.
Adoption14/100
Some open-source and social mention vs Jev, but still obscure to most practitioners a few days after launch.

Vendor claims

~33–40ms per single question on T4 (author self-reported)[Vendor claim — not independently verified]
typed-decisions fine-tuned checkpoint accuracy ~0.77 (author split)[Vendor claim — not independently verified]
zero-shot typed-decisions ~0.36 (near chance; below majority ~0.46)[Vendor claim — not independently verified]

News covering this model

Laya vs Jev — T-Rex arena demo on Apple Silicon

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

Laya is a full open System One–style decision model — ModernBERT-large fine-tuned with a decision head (~421M parameters), Apache-2.0 — not a LoRA adapter and not a TypeSafe product.

You load a checkpoint, pass shared state plus typed questions (choice, score, noul → catalog boolean), and get answers with probabilities in one forward pass. Packaging: PyPI laya, Hugging Face convaiinnovations/laya, source github.com/NandhaKishorM/laya. Family includes English, multilingual (mmBERT-base ~322M), and typed-decisions variants.

Score update (3 Oct 2026). Maturity was recomputed with the weighted formula used for every other page (availability 30% + docs 25% + integrations 25% + support 20%): 9, 7, 6 and 1 give 6.2. The earlier 7.0 came from before that formula. Adding llama.cpp support is already counted in the 6.

Limits (read before you ship)

Zero-shot typed-decisions is weak (~0.36) — near chance; the ~0.77 figure is a checkpoint fine-tuned on that benchmark split, not generic zero-shot System One transfer.
Overconfident until temperature refit — author reports ECE improving a lot after fitting temperature on your data (card: ~0.466 → ~0.081). Do not ship raw probs as calibrated without your own fit.
Comps vs Jev on the card use different samples — Jev was not measured by the Laya authors on the same suite; any “above Jev” claim is not independently verified here.
Choice option budget is limited (author: ≲ ~20 options by token budget).
Independent lab note (unverified): @michhachula (19 Sep 2026) ran Laya locally (~29 ms / short question on M4 Pro) inside a 38-model router: passed 1/5 hard constraints; reversing option order changed 15/16 picks; wrong picks at >98% confidence; ~192-token question budget trimmed model details. Self-reported — not a ModelSystem.One eval. Reinforces: do not treat as drop-in Jev without your own suite.
Latency ~33–40ms / 1q on T4 is self-reported; CPU / other GPUs will differ.

Community framing (unverified)

A Brazilian community post framed Laya as “open source do Jev / already above it” (@0xCVYH, 19 Sep 2026). Treat that as UnverifiedClaim / marketing awareness, not a ModelSystem.One benchmark. Useful only as an adoption signal (~60k impressions on that post). Update: JevBench v1.4.2 (#36) definitively debunked the “above Jev” claims.

JevBench v1.4.2

Metric	Score	Rank
Overall	30.3	#36
Intelligence	38.2	-
Calibration	62.1	-
Speed	72.8	-
Cost	71.4	-

The independent benchmark confirms the catalog’s earlier caution: Laya’s zero-shot typed-decisions score (~0.36) translates to weak JevBench performance. The “above Jev” community claims were marketing noise, not reality. Laya remains valuable for its packaging and fine-tuning potential — not as a drop-in Jev replacement.

Decision Index 0.2.1 (official board)

On the official Decision Index 0.2.1 board (snapshot 28 Sep 2026), Laya scores 6.04, 60th of 70 open entries. The maintainers measured a median of 5.8 ms per request on an RTX PRO 6000 and an ECE of 0.140. Jev scores 57.91 on the same board. The index measures zero-shot answers on 38 kinds of question the model has never seen, and this is the base checkpoint with no tuning. That matches the README’s own warning, and it is how Fabio Akita’s post (3 Oct 2026, in Portuguese) reads it: weak as a zero-shot generalist, useful when tuned on a stable task with your own labels. No score change; the capability of 5.0 already reflects this.

Adjacent (choice-only encoder)

MaziyarPanahi/ModernJEV-Decide-Preview is a 149.6M ModernBERT cross-encoder for choice ranking (next-action / tool selection). Author numbers: next-action 73.4%, tool selection 62.2%, When2Call 34.4% vs 35.5% majority baseline (below chance-of-majority). Preview, choice-only — adjacent, not catalog. See docs/04.

Project	Role
virajbhartiya/laya-vs-jev	T-Rex arena: local Laya via MLX vs hosted Jev; includes the laya-mlx Apple Silicon runtime. Assisted play uses planner + safety overrides — not a raw head-to-head skill bench. Latency tables in-repo = UnverifiedClaim
mizorewww/laya-mlx	Upstream MLX port referenced by the arena (HF `aac6fef/laya-*`); independent of ConvAI — not an official release; see Laya-MLX (0.5.0, 20 Oct)

Why it is in the catalog

Public typed decision I/O, open weights + PyPI docs, and a complete encoder+head model (not LoRA). It sits next to Jev (hosted public beta) and Kev (open LoRA prototype) as the stronger open full-model point on the quadrant — with the caveats above on zero-shot and calibration.

JevBench v1.6.1 (Benchmark Heaven, checked 7 Oct 2026)

On the v1.6.1 open-weights board (not comparable with v1.5.7), Laya (ModernBERT-large 421M) is #94 at 0.0 (Intelligence 1.8); Laya typed-decisions #83 at 0.2; Laya multilingual #96 at 0.0. As on v1.5.7, read this as an encoder falling under the board’s Intelligence gate, not as a measured zero accuracy. Decision Index 0.3: Laya 4.43 (rank 107 group). See news. No catalog score change.

JevBench v1.5.7 (earlier edition, checked 5 Oct 2026; not comparable with v1.6)

Laya rows on v1.5.7 show official score 0.0 (ranks #91–#111 depending on variant). Treat as a board gap, not a measured zero accuracy. See news. No catalog score change.

Community fine-tune: Laya TR (Turkish, 3 Oct 2026)

hayriyigit/laya-tr fine-tunes convaiinnovations/laya-multilingual (322M) for 2 epochs on 285,363 Turkish decisions: machine-translated bev-decision-150K, its English originals, and native Turkish MASSIVE, XNLI and movie reviews. Per-type temperatures were fitted on unseen Turkish decisions.

Author numbers (UnverifiedClaim): on 89,838 held-out decisions, accuracy 0.432 → 0.681 and ECE 0.277 → 0.070 vs the base; on the machine-translated jev-bench-tr (22,773 decisions), accuracy 0.402 → 0.490 and ECE 0.295 → 0.171.

Caveats: license “other / mixed-training-data”: the base is Apache-2.0 but the training-data licenses differ, and the card says to check before commercial use. The usage snippet points to a private laya-tr-test repo. A banking variant, laya-tr-karar, went up on 5 Oct. Not on any board. Scores on this page are unchanged.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board	Row	Result
Decision Index 0.3.2 (16 Oct 19:00 UTC)	Laya	Full 4.43 (#164, tie group 49); public 6.19 (#167); same-skills 9.08; new-domains 4.05
JevBench open weights (live 17 Oct)	Laya	Capability 32.5; Intelligence 1.8; calibration 63.2; p50 0.825 s; $0.0032/1k
JevBench open weights (live 17 Oct)	Laya typed-decisions	Capability 44.2; Intelligence 5.5; calibration 83.0; p50 0.864 s; $0.0038/1k
JevBench open weights (live 17 Oct)	Laya multilingual	Capability 17.0; Intelligence 0.3; calibration 33.7; p50 0.358 s; $0.0039/1k

Board-status statements on this page to recheck against the table (may be superseded):

“Update: JevBench v1.4.2 (#36) definitively debunked the “above Jev” claims.”
“On the official Decision Index 0.2.1 board (snapshot 28 Sep 2026), Laya scores 6.04, 60th of 70 open entries.”

Score update (audit, 17 Oct 2026). Capability 5.0 → 5.5, first full working (the earlier score had no stored breakdown) = types 10 × 0.30 + calibration 4 × 0.25 + latency 5 × 0.25 + accuracy 1 × 0.20 = 5.45 → 5.5. Types from decisionTypes; calibration from the JevBench axis 63.2 (≥90 → 7, 85–89.9 → 6, 75–84.9 → 5, 60–74.9 → 4, below 60 → 3; no author ECE counted); latency from the board’s self-hosted p50 0.825 s without p99 (≤0.25 s → 7, ≤0.5 s → 6, ≤1 s → 5, above → 3); accuracy Decision Index Full 4.43 ÷ 60.11 × 8 = 0.6, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (convaiinnovations/laya): 50,888 downloads, 5,668 likes, licence tag apache-2.0, last modified 2026-10-03.

Community ONNX for WebGPU (24 Oct 2026)

publicado (OK do Fabricio 25/10 BRT).

ningsuhen/laya-typed-decisions-onnx (Apache-2.0, 24 Oct 19:52 BRT) is an unofficial fp16 ONNX export of convaiinnovations/laya-typed-decisions for ONNX Runtime Web on WebGPU, with the NaN guards rewritten so they run on the GPU. No retraining. Author parity checks (UnverifiedClaim): same top option as native fp32 on 47/47 browser passes (max probability difference 0.021); 0 flips in 3,082 decisions on a native CPU subset; about 52 ms per forward on one Apple GPU. Tokenization is left to the caller with the upstream laya package. Scores unchanged.