laya
Classifies text into typed answers in one pass, 33ms, no generation to hallucinate.
Answers typed questions (choice, score, or yes/no) about any input — email, ticket, JSON — in a single forward pass, no text generation, so there's nothing to parse or hallucinate: 33ms for one question, 7.2ms batched, on a T4 GPU. A built-in Router auto-detects the language and script, then dispatches to one of three checkpoints — English-only ModernBERT-large (421M params) or the 100+-language mmBERT-base — and returns the reason for that choice with every result, e.g. 'non-Latin script (devanagari); the English checkpoint cannot read it.' Install is a single pip install laya.

What holds up
- +No text generation — typed choice/score/boolean answers only, so there's nothing to hallucinate.
- +33ms per question (7.2ms batched on a T4), fast enough to run inline on live requests.
- +Every answer includes why the router picked that checkpoint, not just the result.
- +Covers 100+ languages via a second checkpoint, not just English.
Mind the limits
- −0% accuracy despite 95% confidence on Khmer in their own benchmark — multilingual coverage isn't even.
- −Self-hosted Python package, not a hosted API — you manage the GPU/inference yourself.
Picked here? Add the badge
Put it on your site. It links back to this page. Free for every listed product.
<a href="https://stillworks.dev/products/p/laya/"><img src="https://stillworks.dev/badge/laya.svg" alt="Picked by StillWorks" width="230" height="54"></a>