Fuel Labs
ReportFuel Labs is an Indian open-source AI lab founded by Prabhurajhan R. (Prajhan) and Hariharan M. S., both based in India. We build efficient open language models designed to run locally. We pretrained HobbyLM-1B, a sparse mixture-of-experts model with 1.04B total and about 305M active parameters, from scratch on 100B tokens for under $2,000 using GPU credits from Jarvis Labs, and released base and instruct models, documentation and an arXiv paper.
HobbyLM
HobbyLM is an open, from-scratch sparse mixture-of-experts language model family built in India. HobbyLM-1B has 1.04B total and about 305M active parameters (64 routed experts, top-8, plus 1 shared expert), pretrained on 100B tokens for under $2,000. Base and instruct models, documentation and a public paper (https://arxiv.org/abs/2610.34582) document what worked and what broke. It is a low-cost, reproducible example for students, researchers and small teams, and it runs locally: a computer-use demo on a MacBook Air completed 90 of 125 sealed tasks with local inference. About 990 Hugging Face downloads so far. Next open release: HobbyLM-Embed, a compact embedding model for local retrieval.
HobbyLM open development grant
One-time grant over 12 months for maintainer time on HobbyLM. M1 (month 3): improved docs, a reproducible evaluation harness and simple local-run guides for laptops and Macs. M2 (month 6): open release of HobbyLM-Embed weights with published evaluation results, plus a talk or blog post for the FOSS United community. M3 (month 9): quantized local builds, contributor guide and issue triage. M4 (month 12): next open release and a final public report. Quarterly updates on the FOSS United forum.