Laya open-sources a 33ms multilingual typed-decision engine

Laya, a new open-source engine on GitHub, evaluates typed questions in a single forward pass at 33 ms on a Tesla T4, rather than generating text. Its README pitches it as a "multilingual, non-autoregressive System 1 decision engine" handling choice, score, and noul (yes/no/unknown/likely) primitives over "any state (text, email, ticket or JSON document)."

The project ships three checkpoints, an English 512-token model, a multilingual 1024-token model covering 100+ languages, and a fine-tuned typed-decisions variant, all built on ModernBERT-large and mmBERT-base. A router "detects the script in <0.5 ms pure Python before the forward pass" and dispatches to the appropriate model. Against a system called Jev, the repo claims Laya is "roughly 6-7x faster" (32.8 ms vs. 236 to 276 ms) and its typed-decisions checkpoint reports 0.766 accuracy versus Jev's published 0.727. On the MASSIVE intent task, the English model reaches 0.783 in its own language but drops to 0.306 across 13 others; the multilingual variant scores 0.451 on the same set.

The repo is candid about failure modes. "The base checkpoints are near chance on typed-decisions zero-shot -- 0.362 and 0.352 against a 0.318 random baseline," and "ordinal `score` questions are the weakest primitive (SST-5 0.372)." The English checkpoint "collapses on non-Latin scripts": Khmer scores 0.000 at 95.2% confidence, and only 23 of 51 languages clear a 3x-random threshold on that model. Post-hoc temperature scaling moves mean expected calibration error from 0.466 to 0.081.

Weights sit on Hugging Face under convaiinnovations/laya, installation is pip install laya, and a fine-tuning notebook targets 2xT4 GPUs. The repo has already surfaced in feeds from two researchers on our tracker.