Laya

// Overview

Laya is a family of small open decision models from Convai Innovations, published on Hugging Face on September 18, 2026 under Apache-2.0. It never generates text: you give it a state (text, an email, a ticket or JSON) and typed questions — `choice` picks an option, `score` places a value on an ordinal scale, `noul` returns the probability that a statement is true — and every answer comes back as a typed value with probabilities, all questions in a single forward pass.

The family has three checkpoints in one Hugging Face repo. `laya` (421M) is built on ModernBERT-large for English with a 512-token budget; `laya-multilingual` (322M) is built on mmBERT-base for 100+ languages with 1,024 tokens, extendable to 8,192; `laya-typed-decisions` (421M) is fine-tuned on the four workflows of the typed-decisions benchmark. The `laya` Python package ships a Router that detects the script and language of each input and sends it to the right checkpoint.

Training uses what the authors call RLCD (Reinforcement Learning for Calibrated Decisions): the reward is a strictly proper scoring rule, so the policy earns the most by reporting honest probabilities. The model card still says the checkpoints ship over-confident and recommends refitting a temperature on your own data, which moves mean ECE from 0.466 to 0.081 for `laya`.

On a Tesla T4, `laya-multilingual` answers one question in 32.8 ms and ten in 72.3 ms. The card compares Laya against third-party published figures for TypeSafe's hosted Jev 1.13.0: 0.766 against 0.727 on typed-decisions and 0.950 against 0.910 on AG News, while Jev leads on Banking77 (0.870 against 0.425). It is explicit that the base checkpoints are near chance on typed-decisions zero-shot (0.362 English, 0.352 multilingual) and that the 0.766 belongs to the checkpoint fine-tuned on that benchmark's training split.

The package includes a Jev-compatible HTTP server (`laya-serve`, `POST /v1/systemone`), an MCP server, LangChain and LangGraph integrations, an ONNX path and a Kaggle fine-tuning notebook.

// Benchmarks

This model's scores

1. typed-decisions accuracy (laya-typed-decisions)76.6%
2. AG News accuracy (routed)95%
3. XNLI English accuracy (laya)86%
4. MASSIVE intent, English accuracy (laya)78.3%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• One forward pass answers every question in a call: 32.8 ms for one question and 72.3 ms for ten on a T4 (laya-multilingual)
• Fine-tuned laya-typed-decisions scores 0.766 on typed-decisions, above the 0.735 teacher self-agreement ceiling
• Router picks the English or multilingual checkpoint per request; 45 of 51 tested languages score above 3x random
• Apache-2.0 weights with a Jev-compatible HTTP server, so existing TypeSafe clients work by changing the base URL
• Small enough for CPU serving (193–464 ms per request with preloaded checkpoints) and Apple GPUs

// Best for

• Reach for it to route, triage or score tickets, emails and chat messages where a full LLM call per decision is too slow or costly.
• Reach for it for guardrail and moderation checks that should return a probability you can threshold.
• Reach for it as a fast base to fine-tune on labelled decisions from your own domain, which is where the model card says accuracy jumps.
• Look elsewhere for zero-shot decisions or choice questions with more than about 20 options, where the card reports near-chance and Banking77-style drops.

// How to access