github.com/OpenFrontierLabs/openfrontier
Open Frontier Labs is a community-owned open-weights lab. It is what OpenAI should have been, with a pure community vibe: the work is public, the people who contribute own it together, and nobody sits between the community and the model.
We are just starting. Nothing has been trained yet. There are no models, results or benchmarks to point at. What exists today is a plan and the scaffolding to build it.
Open Infrastructure. Trainers, data pipelines, the node agent and dashboards, all in public repos. If we need a tool to train a model, you can read it, run it and fix it.
Distributed Compute. We intend to train on compute the community pools: pretraining and asynchronous reinforcement learning spread across GPUs that people donate or contribute. The techniques exist in public research; putting them to work as one open effort is the part we are trying to do. Whether it works at useful scale is something we will find out, in public.
Intelligence Research and Exploration. Our own base models, evals and post-training, with negative results published alongside the good ones. We aim for the best model we can make at each size, not parity with closed frontier labs.
Data recipes, code, configs, training logs, intermediate checkpoints and final weights, released as they happen. Code and weights are licensed under Apache-2.0.
- Sign up: openfrontier-labs.com/join, then say hi in Discord
#introductions. - Pick a task: good first issues, each with context and a definition of done. How to contribute.
- See the plan: the public roadmap and the proposed milestone ladder.
- Ask and propose: Discussions for Q&A, ideas, show and tell and RFCs.
- Contribute compute: what works today is in docs/join-a-node.md. To offer hardware, use the compute offer form.
- Read the manifesto: MANIFESTO.md.
Start at OpenFrontierLabs/openfrontier.