Author
Convai Innovations (Hugging Face)
1 reading card from 1 book · 2026.
1 card
Laya — model card · 2026
Laya speaks Jev's contract on your own hardware, but it is a base to specialise, not a zero-shot engine.
Convai Innovations published Laya on 18 September 2026, three days after Jev, with three checkpoints on Hugging Face: `laya` (ModernBERT-large, 421M parameters, 512 context, English), `laya-multilingual` (mmBERT-base, 322M, 1024 context, 100+ languages) and `laya-typed-decisions` (421M, fine-tuned on decision workflows). The architecture explains the speed: a bidirectional encoder that scores the options in one pass, not a decoder writing token by token. Measured by the authors on a Tesla T4: 39.5 ms for one question on the English checkpoint, 32.8 ms on the multilingual one, 7.2 ms per question in batches of ten; on CPU, 193–464 ms with the model preloaded; about 2 GB of GPU memory per checkpoint. The contract is deliberately the same: the `questions` dictionary with `choice`, `score`, `noul` and a result with `choice`, `probabilities`, `confidence`. A `Router` detects the text's script in under 0.5 ms and picks the checkpoint before the forward pass — necessary, because the English model "collapses" outside English without lowering its confidence. The limits, in the authors' own words. The base checkpoints are near chance on typed decisions zero-shot (0.362 against 0.318 random and 0.461 majority class); the 0.766 figure that beats Jev (0.727) comes from the checkpoint fine-tuned on that benchmark's own training split — "a fast base to specialise, not a zero-shot decision engine". Past 20 options the token budget per label thins out (0.425 against Jev's 0.870 on 77 labels). And it ships over-confident: ECE 0.466 before calibration, 0.081 after.
“Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.”