Solo dev open-sources Laya: 421M non-autoregressive decision model hitting 35ms forward passes
Nandakishor_ml · reddit · 2026-09-18
Reddit user Nandakishorml released Laya, an open-source decision model trained on a single RTX 6000 Pro (96 GB VRAM) as the generalized follow-up to his earlier "Jev architecture" project.
- Architecture: a 421M-parameter non-autoregressive decision model pairing a bidirectional ModernBERT-large encoder with a scratch Transformer head that scores [MASK] option markers, resolving typed schemas in a single 35 ms forward pass.
- Data: 25,000+ fully human-annotated real-world examples spanning intent routing, fact-checking, moderation consensus, prompt guardrails, rubric scoring, and multi-turn conversation trajectories — no synthetic shortcuts.
- Method: unofficial RLCD, a policy-gradient RL approach optimizing against strictly proper scoring rules so maximum reward requires true, mathematically calibrated probabilities.
He claims it surpasses all prior Jev benchmarks and published an HF Space demo, GitHub repo, and HF weights, while noting the model needs further refinement.
More from Research
- Fei-Fei Li: 540 million years of vision evolution drove the development of intelligence — rohanpaul_ai · 2026-09-18
- Researcher slams conference rule banning post-acceptance paper revisions at ICSE — tianyin_xu · 2026-09-18
- Visual General Intelligence white paper authors to present at Video Model Journal Club — HirokatuKataoka · 2026-09-18
- The subtle KV cache detail most people get wrong: tokens generate before entering the cache — techNmak · 2026-09-18
- You can turn jev back into an autoregressive LLM just by making it output ASCII — multiply_matrix · 2026-09-18
- Follow-up: December paper claims 8.1 Mbps peak throughput, but major physics hurdles remain — BlancheMinerva · 2026-09-18