funding.json directory

laya

laya is an open-source System 1 decision model. It reads a message once and answers typed questions (a choice, a score or yes/no) in a single forward pass, about 33 ms on one T4 GPU, with calibrated probabilities you can branch on. It runs on a laptop, a CPU, a phone or an edge board, so the data never leaves the machine.

Since its release in September 2026: 35,000+ GitHub stars and 3,200+ forks, the front page of Hacker News, #1 trending on Hugging Face with 370+ community models built on it, and 250,000+ PyPI downloads a month. It is built into the open-source projects of 43+ teams, including Microsoft PyRIT, NVIDIA TensorRT-Model-Connect, Tencent tRPC, Spring AI, Mozilla.ai llamafile, Tether QVAC, LiteLLM, Unsloth and Ollama.

laya 0.5 reads images, audio and video, takes 73k tokens of context, says when it should not decide (abstain and unbound flags), and keeps learning from the answers a team confirms, in seconds on a CPU. Weights, training code, benchmarks and notebooks are all open under Apache 2.0.

Funding pays for GPU time to train and openly benchmark checkpoints, Indian and low-resource languages, maintenance of the integrations and ports, and small models for phones and CPUs.

Fund this project

Report listing

Loading report form…

Unverified URL

The funding manifest has not provided proof via wellKnown that this link is associated with it. Learn more (opens in a new tab).