Laya makes AI decisions in 33 milliseconds without generating text

Most LLMs waste time generating full sentences just to say "yes" or "no" — Laya skips text generation entirely.

Developer Nandha Kishor M released Laya, an open-source decision engine that picked up over 20,000 GitHub stars this week. Instead of generating text token by token, Laya evaluates inputs in a single forward pass to return typed choices, scores, or yes/no decisions over text, emails, or JSON documents. It supports 100+ languages, answers single questions in 33 milliseconds on a T4 GPU, and uses an internal router to select the right model checkpoint automatically.

Why it matters: Standard LLM classifiers are slow, expensive, and fail when text output strays from expected schemas. Because Laya doesn't generate text, there is nothing to parse and nothing to hallucinate. It gives you fast, structured decisions for triage and workflows at a fraction of the compute cost.

Here's the gist: Laya provides three checkpoints built on ModernBERT and mmBERT, spanning 322 million to 421 million parameters. You can install it via PyPI (pip install laya) to run single or batched evaluations across languages without managing models manually.

If your application just needs a fast verdict, stop paying LLMs to talk to themselves.