Models › laya family

laya-multilingual

convaiinnovations Apache 2.0 out 36 days ago

322M parameters, text in, labels out.

Verified 2026-10-25 Rank #0 in Decision models · score 0 method v1 Sources 3 Downloads 30d 14.9k · #478 of 619 tracked

Changed this week: 7 new quantized builds on the Hub (45 in all).2026-10-24

Can I run it?0.2GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?14.9kdownloads, 30 days#478 of 619 we track · below the median+12 likes, downloads +19% this week406 likesHugging Face2026-10-24

Worth reading

Picked by rule from 5 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 15 HN pts

Does an open-weight decision model beat a hosted one? Jev vs. Laya

The developer deployed the Apache 2.0 licensed Laya model, which contains about 421 million parameters, locally on a Mac Studio to avoid hosted service constraints.

  • The service fails closed by refusing to download models during inference and requiring Apple's Metal accelerator.
  • The gateway caps Choice at 20 options by default and serializes inference to prevent uncontrolled accelerator load.
astgl.com 2026-09-22 15 HN pts · 7 comments 2.6k words about the laya family family
Latest · 30 days ago

Benchmarking Jev, Laya, and five open models under three stresses

Eight decision models are evaluated as candidate lists grow, option orders change, and wrong answers become harder to distinguish.

  • Decisions returns an exact zero flip rate with fixed order despite having no seed parameter or temperature control.
  • Jev recomputes and agrees with itself after changing input token count by one trailing space.
github.com 2026-09-25 4 HN pts 4.7k words about the laya family family

5 best of 5 tracked

The 2 picks above, plus these 3. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleLaya — 33ms Multilingual System 1 Decision Enginelaya.convaiinnovations.com · 2026-09-19
  2. ArticleLaya, an open decision model that plays Tetris by itselfbrainfunctioncollapse.com · 2026-09-22 · 4 HN pts
  3. ArticleFree hosted API for Laya, the open-weight decision modelvercel.com · 2026-10-01 · 2 HN pts

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 45 quantized builds and 16 Spaces.

  • As a service0no API provider listed
  • On own hardware45quantized builds
  • Built on8484 fine-tunes · 0 adapters · 0 merges
  • In apps16Hugging Face Spaces

Top Spaces running it: mayafree/verifier-playground ♥ 39 · abhishekbhakat/laya-multilingual-demo ♥ 7 · hugging-apps/laya-hebrew ♥ 1 · walidhadri/playground-decision-models ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 0.2 GB at INT4, 0.4 GB at FP8, 0.8 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.2 GB of weights (estimate; KV cache and activations extra)INT40.2 GBFP8: about 0.4 GB of weights (estimate; KV cache and activations extra)FP80.4 GBBF16: about 0.8 GB of weights (estimate; KV cache and activations extra)BF160.8 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 0.2 GB of weights (estimate; KV cache and activations extra)INT40.2 GBFP8: about 0.4 GB of weights (estimate; KV cache and activations extra)FP80.4 GBBF16: about 0.8 GB of weights (estimate; KV cache and activations extra)BF160.8 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Coverage peaked at launch and is still arriving

Teal is the weights (2026-09-19), amber triangles are 5 dated articles (each links to it), grey ticks are repo commits.

What people built on it

129 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

129 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

45 quantized (Hugging Face model tree)84 fine-tunes (Hugging Face model tree)45 quantized84 fine-tunes
45 quantized (Hugging Face model tree)84 fine-tunes (Hugging Face model tree)45 quantized84 fine-tunes

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

9 rows

Benchmarks 9 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (9)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported Arabic0.4intent classificationmodel author2026-10-07
author-reported Azerbaijani0.3intent classificationmodel author2026-10-07
author-reported Bengali0.29intent classificationmodel author2026-10-07
author-reported Hindi0.39intent classificationmodel author2026-10-07
author-reported Korean0.49intent classificationmodel author2026-10-07
author-reported Turkish0.44intent classificationmodel author2026-10-07
author-reported languages clearing 3x random45 / 51intent classification with 20 options (all 51 MASSIVE languages)model author2026-10-07
author-reported macro ECE0.39intent classification with 20 options (all 51 MASSIVE languages)model author2026-10-07
author-reported macro accuracy0.37intent classification with 20 options (all 51 MASSIVE languages)model author2026-10-07

Use-case evidence

4 use cases

What it is good at 4 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

17 fields

Listed in Decision models

Facts

Pipeline tagtext-classification
Librarytransformers
ArchitecturesLayaTypedDecisions
Model typelaya
Params by dtypeF32: 3 · F16: 321,908,995
Size bandS
MoEno
Languagesmultilingual, en, de, fr, es, pt, it, nl, sv, da, nb, ru, pl, tr, ar, he, fa, ur, hi, bn, ta, te, kn, ml, th, vi, id, ms, tl, ja, ko, zh, el, hu, fi, ro, sq, sl, sw, af, cy, am, hy, ka, km, my, mn, lv, is, az, jv
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for laya-multilingual: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)0.8 GB · Edge (<4 GB)
FP8 (1 byte/param)0.4 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)0.2 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

Decision facts

catalog

Decision-model catalog curated

How it was madehead + encoder trained with RLCD-named objective, 4 epochs ~4.97 h
Lineage bucketF - fine-tune of an open model (public/synthetic data)
Lineage evidencemodel card
Base modelmmBERT-base
Hosted price (input)self-hosted/M
Sourcehttps://huggingface.co/convaiinnovations/laya-multilingual

All variants

40 variants

Lineage

Original model Family: convaiinnovations/laya 129 derivatives on the Hub

quantized: 45finetune: 84

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
mys/laya-multilingual-GGUF quantized GGUF — — apache-2.0 mys 15 11.6k
Weidows/laya-multilingual-GGUF quantized GGUF — — apache-2.0 Weidows 6 4.6k
fr0stbit3/laya-multilingual-gguf quantized gguf — — apache-2.0 fr0stbit3 1 1.3k
aac6fef/laya-multilingual-coreml-ane quantized — — — apache-2.0 aac6fef 10 657
meshllm/laya-multilingual-F16-GGUF quantized GGUF — — apache-2.0 meshllm 1 831
onnx-community/laya-multilingual-ONNX quantized — — — apache-2.0 onnx-community 4 651
ldov/laya-multilingual-GGUF quantized GGUF — — apache-2.0 ldov 0 823
aac6fef/laya-multilingual-coreml quantized — — — apache-2.0 aac6fef 5 568
andreilungeanu/laya-multilingual-GGUF quantized GGUF — — apache-2.0 andreilungeanu 0 792
aac6fef/laya-multilingual-mlx finetune mlx 322M S apache-2.0 aac6fef 15 0
mccoysc/xDecision finetune — — — apache-2.0 mccoysc 1 483
Wouze/laya-ara-rag finetune — 322M S apache-2.0 Wouze 3 325
telepatia-ai/laya-pt-es-typed finetune — 322M S apache-2.0 telepatia-ai 4 214
Wouze/laya-ara finetune — 322M S other Wouze 3 243
Beko2210/statim-decide-multilingual-base quantized GGUF — — other Beko2210 1 324
letrthang/aimtas-laya-multilingual-GGUF quantized GGUF — — apache-2.0 letrthang 0 370
Ramg77/laya-sentiment-multilingual finetune — 322M S apache-2.0 Ramg77 0 315
hayriyigit/laya-tr finetune — 322M S other hayriyigit 5 38
ti3x-m/laya-multilingual-onnx quantized — — — apache-2.0 ti3x-m 0 242
OidoStudio/laya-mm-guard-v3-gguf quantized gguf — — apache-2.0 OidoStudio 1 171
russian-oracle/laya-multilingual-GGUF quantized GGUF — — apache-2.0 russian-oracle 0 212
hayriyigit/laya-tr-karar finetune — 322M S other hayriyigit 3 60
mizchi/laya-multilingual-onnx quantized — — — apache-2.0 mizchi 4 0
alfred361/laya-multilingual-typed-decisions finetune — 322M S apache-2.0 alfred361 4 0
Suarify/laya-multilingual-onnx quantized — — — apache-2.0 Suarify 0 184
kazalbrur/siddhanto-bangla-english-banglish finetune — 322M S apache-2.0 kazalbrur 1 70
felhen-ai/saracura-ptbr-v0 finetune — 322M S apache-2.0 felhen-ai 1 58
smolnikov/kivok-0.3b finetune — 322M S apache-2.0 smolnikov 1 52
Weidows/laya-multilingual-FP8 quantized -FP8 — — apache-2.0 Weidows 2 0
soyelmismo/laya-multilingual-onnx quantized — — — apache-2.0 soyelmismo 2 0
Emerald7664/laya-multilingual-onnx quantized — — — apache-2.0 Emerald7664 2 0
telepatia-ai/hertz-1 finetune — — — other telepatia-ai 2 0
amiralifirouzi/dana-laya-0.3b-persian finetune — 322M S apache-2.0 amiralifirouzi 1 41
ParkJH09/ownd-laya-search finetune — 322M S — ParkJH09 1 39
JonusNattapong/jev-my-bro-th1200 finetune — 322M S apache-2.0 JonusNattapong 1 38
ahmedandaloes/laya-multilingual-CoreAI quantized — — — — ahmedandaloes 0 77
snowman6/Miya-0.2 finetune — — — apache-2.0 snowman6 1 22
rehman-ali/laya-tetris finetune — 322M S apache-2.0 rehman-ali 1 21
MahdiAlikhah/laya-persian-sentiment finetune — 322M S apache-2.0 MahdiAlikhah 1 18
goktugoguz/laya-multilingual-arc finetune — 322M S apache-2.0 goktugoguz 1 17

Repository changelog

14 commits
changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 14 commits

DateCommitSHA
2026-10-03 chore: root config.json so the Hub tracks downloads 1720e3e
2026-09-24 README: recommend max_len=8192, laya 0.3.20 notes e4e9ddf
2026-09-24 README: measured long-document table, laya 0.3.19 notes 358bda8
2026-09-24 README: 8,192-token long documents on laya-multilingual, laya 0.3.18 notes af8d1ac
2026-09-24 README: laya 0.3.14 release notes and docs site 453577a
2026-09-24 README: laya 0.3.13 release notes a2637a2
2026-09-24 README: laya 0.3.12 release notes 33db9ee
2026-09-23 README: laya 0.3.11 release notes 82d57fc
2026-09-23 Model card: laya 0.3.7 routing and limits b4a904d
2026-09-19 Add the Laya mark to the card; document Router.preload / Router.attach 052592a
2026-09-19 chore: point encoder at the upstream model id instead of a local build path adba930
2026-09-19 docs: measured benchmarks, family hub links, Router usage, honest limits f6f7407
2026-09-19 Initial release of Laya Multilingual (mmBERT-base, 100+ languages, RLCD) 4bb4d65
2026-09-19 initial commit 3357320

Articles and coverage

5 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis5

astgl.com2026-09-22about the laya familyHN 15 pts · 7 comments2,557 words
The developer deployed the Apache 2.0 licensed Laya model, which contains about 421 million parameters, locally on a Mac Studio to avoid hosted service constraints. On a held-out replay of 40 decisions, this local setup achieved a 10-percentage-point gain in acceptable routing decisions compared to a deterministic baseline, though p95 latency reached 466.69 ms under load. The system uses a custom runtime to render questions and options into a single token sequence, eliminating the need to parse unstructured text.written, reviewed, cited
  • The service fails closed by refusing to download models during inference and requiring Apple's Metal accelerator.
  • The gateway caps Choice at 20 options by default and serializes inference to prevent uncontrolled accelerator load.
  • Laya's publisher warns the model is overconfident and accuracy falls as the option list grows.
github.com2026-09-25authorityabout the laya familyHN 4 pts · 0 comments4,661 words
Eight decision models are evaluated as candidate lists grow, option orders change, and wrong answers become harder to distinguish. OpenAI's Decisions API and TypeSafe's Jev lead in accuracy, with Decisions costing $0.029 per 1,000 decisions and Jev costing $0.035 per 1,000 decisions. When option order changes, Decisions changes its answer on one in three decisions while Jev changes on one in seven. Holding the order fixed reveals Decisions never changes its answer at all, whereas Jev changes on 4.3% of decisions due to internal non-determinism.written, reviewed, cited
  • Decisions returns an exact zero flip rate with fixed order despite having no seed parameter or temperature control.
  • Jev recomputes and agrees with itself after changing input token count by one trailing space.
  • Decisions loses 27.9% accuracy on near-neighbour distractors for request routing, while Jev loses only 9.0%.
laya.convaiinnovations.com2026-09-19about the laya family1,940 words
Laya is a non-autoregressive decision engine that executes in 32.8 milliseconds on a single GPU with full support for over 100 languages. The model family includes specialized checkpoints like convaiinnovations/laya-multilingual which runs at 7.2 ms per question when batched, eliminating API subscription costs via 100% open-source Apache 2.0 weights. A built-in Router with preload=True inspects Unicode scripts to route requests to the correct checkpoint, ensuring sub-35ms response times without cold-swap penalties.written, reviewed, cited
  • English checkpoint accuracy drops to 0.000 for Khmer while reporting 0.952 mean confidence, proving confidence gating fails for non-Latin scripts.
  • Choice questions degrade significantly with over 20 options, as Banking77 testing showed Laya scoring 0.425 against Jev's 0.870.
  • Base weights ship with raw temperature logits, requiring domain-specific scalar fitting to cut expected calibration error from 0.466 to 0.081.
brainfunctioncollapse.com2026-09-22about the laya familyHN 4 pts · 1 comments1,015 words
Laya is a 322M parameter open-source model that routes tasks and answers typed questions with probabilities in 21 ms on local hardware. It requires 650 MB of disk space and a 2.3 GB download, loading in about 90 seconds to achieve a decision rate of around 30 times per second. Licensed under Apache-2.0, it offers a free, offline alternative to the hosted TypeSafe Jev, ensuring data never leaves the machine.written, reviewed, cited
  • The multilingual checkpoint ships uncalibrated, requiring a language hint to fix routing for non-native scripts.
  • Out of the box Laya is strongest at routing and yes/no checks; graded scores improve a lot with fine-tuning.
  • The first start downloads 2.3 GB of open weights and takes about 90 seconds to load them.
vercel.com2026-10-01about the laya familyHN 2 pts · 1 comments134 words

Similar models

3 models

Similar models

  • sokudan-ja-310m · 315M
    same size band (S), same task (text-classification), permissive license
  • laya-typed-decisions · 421M
    same size band (S), same task (text-classification), permissive license
  • lavoir · 422M
    same size band (S), same task (text-classification)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:43 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:18 UTC · Last verified 2026-10-25

Follow laya-multilingual

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.