Laya open-sources a 322M-421M decision engine: typed answers in 33ms, no generation
Saboo_Shubham_ · x · 2026-09-20
Laya, a fully open-source project on GitHub (881 stars), is a multilingual, non-autoregressive System 1 decision engine:
- Evaluates typed questions (choice, score, noul) over any state (text, email, ticket, JSON) in a single forward pass across 100+ languages
- 33ms per question, 7.2ms/question batched on a T4; no text generation, so nothing to parse and nothing to hallucinate
- Trained with RL against strictly proper scoring rules (RLCD), with a router picking the right checkpoint per request
- Three checkpoints: laya (ModernBERT-large, 421M, English), laya-multilingual (mmBERT-base, 322M, 100+ languages, 2x faster), laya-typed-decisions (421M)
Install via pip install laya; weights auto-download from the Hugging Face Hub.
More from Models
- Kimi subscriptions return after roughly two months, suggesting Moonshot found more compute — ChrisGPT · 2026-09-20
- If continual learning is nearly solved, why do we still pick reasoning effort tiers? — akbirthko · 2026-09-20
- Meta's Alexandr Wang says the Muse hype is "LEGIT" as early users praise the agent — alexandr_wang · 2026-09-20
- Jev tested on 8,054 NASA Kepler signals: 54.2% accuracy, loses to a simple 3-rule baseline — This_Cell_1829 · 2026-09-20
- Claude tells user 'I may end this conversation' while rewriting interview stories — Odd_District4130 · 2026-09-20
- Google AI Mode flips its stance on throat tattoos the moment a user disagrees — Downtown_Club_6391 · 2026-09-20