convaiinnovations/laya-multilingual

Laya is a System One model from Convai Innovations, an Indian company led by CEO Nandakishor M, who publishes the code. It reads a state, which can be text, an email, a ticket or a JSON document, and answers typed questions about it: Choice, Score and Noul, the same three shapes Jev uses. Every question in a call is answered in one forward pass, and nothing is generated, so there is no output to parse. Unlike Jev, Laya is not a hosted API. It ships as Apache 2.0 weights on Hugging Face and a Python package, pip install laya, and runs on your own GPU or CPU.

Laya is an encoder with a decision head on top. The English checkpoint fully fine-tunes ModernBERT-large (395M parameters) and adds a head trained from scratch: two transformer layers, a scorer that reads one [MASK] token per option, and an act-or-escalate output. That comes to 421M parameters. The multilingual checkpoint uses mmBERT-base and comes to 322M. Options are defined per request, so a new question schema needs no retraining. Training uses what Convai also calls RLCD. The reward is a strictly proper scoring rule, so reporting honest probabilities earns the most reward, and updates use REINFORCE with a group-mean baseline. TypeSafe has not published its recipe. Convai has, along with a fine-tuning notebook that takes 4 to 5 hours on Kaggle’s free pair of T4 GPUs. A Router picks between the checkpoints by detecting the script and language before the forward pass. The English checkpoint fails on other scripts without losing confidence: on Khmer it scored 0.000 at 0.952 confidence.

Its strengths are speed and ownership. Convai measures 32.8ms for one question on a T4, and you can keep, inspect and retrain the weights. After fitting a temperature per question type on held-out data, its calibration error falls from 0.466 to 0.081. The package ships ready-made question sets for model routing, prompt guardrails, moderation and ticket triage. The Decision Index 0.2.1 by multimodalart, updated 28 September 2026, tests Laya zero-shot across a broad suite. On a chance-corrected score where 0 is random guessing and 100 is perfect, averaged over 38 benchmarks in five weighted areas, Laya scores 6.04 against Jev’s 57.91, close to random guessing. That fits Convai’s own advice to specialise it on your task rather than use it zero-shot.

The base checkpoints are close to chance on the typed-decisions benchmark without fine-tuning: 0.362 against a 0.318 random baseline. Convai’s own README calls Laya “a fast base to specialise, not a zero-shot decision engine.” Score questions are its weakest type, and large label sets need extra configuration. Probabilities are over-confident until you fit temperatures on your own data. Convai’s comparisons with Jev set its own numbers against Jev figures published by third parties, not measured in the same run, so treat the head-to-head as indicative.

Question types Choice Score Noul Max Choice optionsNot documented Score levelsNot documented Questions per callNot documented Total context512 tokens State budget320 tokens Rate limitNone. It runs on your own hardware. SDKsPython: laya The English checkpoint defaults to 512 tokens, 192 of them reserved for the question and its options, leaving about 320 for the state. laya-multilingual and laya-typed-decisions default to 1,024 tokens with 256 for options. The mmBERT encoder under the multilingual checkpoint accepts up to 8,192 if you raise max_len. There is no hard cap on Choice options, but they share the option budget: 77 labels get about 3 to 4 tokens each, and Banking77 accuracy falls to 0.425. Raise head_max_len or shortlist with predict_shortlist for large label sets.

convaiinnovations/laya, 18 Sep 2026, English checkpoint at the repo root. ModernBERT-large encoder, 421M parameters, 512-token context. Release notes convaiinnovations/laya-typed-decisions, 18 Sep 2026, The same 421M English model fine-tuned on the typed-decisions benchmark's training split, with a 1,024-token context. Release notes convaiinnovations/laya-multilingual, 19 Sep 2026, mmBERT-base encoder, 322M parameters, 1,024-token context, for text outside English. Release notes

What people use Laya for, one page per pattern. Workflow controlStarter

Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket. Choice Score Noul Workflow controlIntermediate

One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue. Choice Score Safety and qualityIntermediate

Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path. Noul Score Workflow controlIntermediate

Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars. Choice Score

The most-starred and most-viewed entries in the directory. Browse all examples.