421M-parameter discriminative model Laya is 7x faster but fails badly on 77-label tasks
maier_ak · x · 2026-09-28
A solo Indian developer released Laya, a 421M-parameter Apache 2.0 model that answers typed questions about text with probabilities in a single forward pass — hitting 1200+ Hacker News points in a day. It responds to TypeSafe AI's Jev (launched Sept 16, $40M raised, CEO Diogo Almeida co-authored InstructGPT), pitched as a "System One model" using choice/score/noul primitives. Andreas Maier argues this marks the return of the discriminative model: fast classifiers are exactly what many production problems need. But benchmarks show the limits — Laya scores 0.950 on AG News at 7x speed, yet loses badly on 77-label Banking77 (0.425 vs 0.870), with its own model card admitting base checkpoints fall below the majority-class baseline zero-shot.
More from Research
- WebArena-Pro: a 300-task multimodal benchmark across 20 self-hosted apps for web agents — McGill-NLP · 2026-11-16
- DySENSE: MIT's GenAI framework synthesizes satellite imagery with dynamic city energy maps — MIT · 2026-11-16
- Continual-ARC: lifelong ARC-AGI stream finds agent harness persistence rarely beats simple caching — unifyai · 2026-11-16
- Most VLAs lose their VLM's vision-language skills during pretraining — DJiafei · 2026-11-16
- Training-free page-sparse decode in llama.cpp: 1.72× faster 128k decoding on an RTX 5090 — topppggg · 2026-11-16
- Training Qwen3-8B to drive a simulated hand spawns a dedicated internal 'pressure' direction — repligate · 2026-11-16