Convai Innovations · 2026-09-18 · major
Laya — an open 421M decision model that replies in 33 milliseconds
Laya is an Apache-2.0 model that answers typed questions about text without generating text. The 421M English checkpoint returns calibrated probabilities in about 33 milliseconds, and a 322M multilingual checkpoint covers 100+ languages.
An Apache-2.0 decision model that returns typed answers with calibrated probabilities in about 33 milliseconds.
Key specs
| GitHub stars | 840 |
|---|---|
| Latency (p50) | 32.8 ms |
Quick facts
| Maker | Convai Innovations |
|---|---|
| License | Apache-2.0 |
| Checkpoints | laya (421M), laya-multilingual (322M), laya-typed-decisions (421M) |
| Base models | ModernBERT-large and mmBERT-base |
| Question types | choice, score, noul |
| Install | pip install laya |
| Availability | Weights on Hugging Face, code on GitHub |
Benchmarks
What is it?
Laya answers typed questions about a piece of text — pick one of these options, score this on a scale, is this true — and hands back probabilities instead of a written reply. Convai Innovations published three Apache-2.0 checkpoints: a 421M English model built on ModernBERT-large, a 322M multilingual one built on mmBERT-base, and a version fine-tuned on typed-decision workflows. The project presents itself as the open counterpart to TypeSafe's Jev.
How does it work?
A single forward pass scores every question at once, with no token-by-token generation, which is where the 32.8 ms median comes from. Training used RLCD, where the reward is a strictly proper scoring rule — the model earns the most by reporting honest probabilities rather than confident-sounding ones. Three primitives cover the work: `choice` picks an option, `score` places a value on an ordinal scale, and `noul` returns P(true).
Why does it matter?
Routing, triage and classification normally cost a full LLM call each. Laya does that work for a fraction of the latency under a licence you can ship, and the calibration training means the confidence number is something you can threshold on. The model card is blunt about the catch: zero-shot accuracy on typed-decisions is 0.362, so Laya is a fast base to fine-tune, not a drop-in decision engine.
Who is it for?
backend and agent engineers
Frequently asked questions
- How much does Laya cost?
- Laya is free. All three checkpoints are published under Apache-2.0 on Hugging Face with the code on GitHub and the package on PyPI, so there is no per-token fee and no hosted API to sign up for. That is the main difference the project draws with TypeSafe's Jev, which is closed and billed per token.
- How does Laya compare to TypeSafe Jev?
- On the typed-decisions benchmark Laya's fine-tuned checkpoint scores 0.766 against Jev 1.13.0's 0.727, with a Brier score of 0.062 that the card puts at 2.4x better than Jev. Median latency is 32.8 ms for Laya versus 236–276 ms for Jev, a roughly 7.8x gap the card credits to non-autoregressive scoring.
- Can Laya be used zero-shot?
- No — Convai Innovations says so directly. Laya scores 0.362 on typed-decisions zero-shot and the multilingual checkpoint 0.352, close to chance. The 0.766 figure belongs to the checkpoint fine-tuned on that benchmark's own training split, and the card describes Laya as a fast base to specialise rather than a ready-made decision engine.
- Which languages does Laya handle?
- The laya-multilingual checkpoint is built on mmBERT-base and covers more than 100 languages; Convai Innovations reports it working on 45 of the 51 languages it tested. The English-only base and typed-decisions checkpoints use ModernBERT-large instead, with 512 and 1024 tokens of context respectively.
Try it
pip install laya