Read the manifesto · Architecture · Roadmap · Discussions · RFCs
Open infrastructure for collective frontier intelligence: fully open models trained on community compute.
Open Frontier Labs is what OpenAI should have been, with a pure community vibe: built in public, meant to be owned by a nonprofit, and trained on compute the community pools. No equity, no token, no closed tier.
Status, plainly: nothing has been trained yet. This repository is scaffolding for the first runs. The CPU smoke tests run; the 1B baseline (Run 0) has a launcher, a config and pinned upstreams, and has not been started on any GPU. There are no results, no released weights, and no public run to join. The foundation described in GOVERNANCE.md has not been formed.
Everything we produce is published, as it happens:
- Data recipes and manifests: dataset ids, licences and revisions in
manifests/. - Code and configs: this repo; every upstream we build on is pinned to an exact commit in
configs/upstreams.toml, and nothing is vendored. - Logs, including failed and diverged runs.
- Every checkpoint, not just the last one.
- The weights, Apache-2.0. Datasets and logs we author are intended to be CC-BY-4.0; data we reuse keeps its upstream terms. No use-restriction licences.
Contributors and donors are credited by name in a public compute ledger and in CREDITS.md.
| Phase | What | Status |
|---|---|---|
| 1. Single-node, fully open pipeline | Run 0: a dense ~1B baseline with the official OLMo-core trainer and the OLMo recipes, on one GPU node. In parallel, an MoE track (about 6.9B total / 1.2B active, 64 experts, top-8) following Hugging Face's Ultimate MoE Training Handbook and Megatron-LM. | Launchers, configs and manifests exist; nothing has run on a GPU. Run 0, MoE |
| 2. Post-training | SFT, DPO and RLVR with open-instruct; async RL with prime-rl and verifiers. | Pinned and documented; launcher not built. Details |
| 3. DiLoCo across islands | DiLoCo outer loops (prime-diloco) over single-node islands: community nodes and donated clusters. | Config and algorithm smoke; a two-machine swarm proof of concept on a toy model; launcher not built. Details, swarm PoC |
Later, once phase 3 exists, and not scheduled: a mid-scale community model of about 10B (Run 1), distributed RL over a swarm of rollout nodes, and a 30 to 70B model (Run 2) on donated clusters plus community nodes. We aim for best-in-class at each size, not frontier parity.
git clone https://github.com/OpenFrontierLabs/openfrontier && cd openfrontier
make setup # uv sync the Python smoke environment (smoke/uv.lock); needs uv and a Rust toolchain
make doctor # GPU, CUDA, network latency, disk, clock, versions; works with no GPU
make smoke # CPU: tiny OLMo-core run, then reference DiLoCo (first run downloads CPU torch once)
make help # every target
cargo run -- run0 launch --dry-run # print the real OLMo-core command for Run 0make check runs cargo fmt --check, clippy, the ledger schema and example, and the config and pin checks; make test runs cargo test, then the swarm tests (smoke/tests, pytest). All targets are one-line wrappers around ofl, the Rust CLI in src/. The two training smokes are Python (smoke/), because they drive the pinned olmo_core and torch trainers; ofl runs them in the smoke/ environment.
What make smoke does and does not prove: it runs the pinned OLMo-core trainer on CPU with a 1.1M-parameter model and checks that the loss falls and a checkpoint is written; and it runs a reference implementation of DiLoCo (not prime-diloco, which needs CUDA) on two CPU workers and checks that the loss falls and the workers agree after each outer sync. It does not exercise the 1B configuration, real data or a GPU.
- Say hi in Discord
#introductions, or sign up at openfrontier-labs.com/join. - Pick a good first issue and comment that you are taking it.
- Set up and check the repo:
make setup && make check && make test. - Commit with sign-off (
git commit -s, see DCO) and open a draft pull request early. - Have GPUs instead of time? Run
make doctorand send a compute offer.
Bigger, hard-to-reverse changes start as an RFC. The milestones from Run 0 to a GPT-3-class run are proposed in RFC 0001 and tracked on the roadmap.
Code, compute and results are all welcome. CONTRIBUTING.md says how; docs/join-a-node.md says plainly what works today (single-node phase 1 runs) and what is planned but not built (joining a shared DiLoCo run).
Builds of ofl compiled with an HTTPS OFL_SENTRY_DSN build-time variable include opt-out crash and unexpected-error reporting via Sentry (panics and top-level OflError failures with exit code != 2, with stack traces, release = ofl@<version>, and environment = development | production). Source builds without OFL_SENTRY_DSN make telemetry a complete no-op; no DSN is committed to this repository, and no prebuilt releases exist yet.
Any opt-out signal disables telemetry unconditionally and cannot be overridden by another variable (for example, OFL_TELEMETRY=1 does not override DO_NOT_TRACK=1):
- Global CLI flag:
ofl --no-telemetry <command>(also recognized before--onofl swarm/ofl node) - Environment variable:
OFL_TELEMETRY=0(also acceptsfalse,off,no; case-insensitive) - Standard opt-out:
DO_NOT_TRACK=1(also acceptstrue,on,yes; case-insensitive)
You can also compile ofl without the sentry crate at all (telemetry is the only default Cargo feature): cargo build --no-default-features.
Privacy guarantees (send_default_pii = false, no contexts OS/device/hostname fingerprint, no traces, no session tracking): before any event leaves the machine, before_send strips user, request, server_name, device.name, stack frame.vars, hostnames, IPv4 and IPv6 addresses, Sentry DSNs, Bearer/Basic tokens, secret-shaped KEY=VALUE assignments, working directory paths (<cwd>), secret CLI flags, sensitive metadata keys, and replaces $HOME and local usernames with ~. The ofl Sentry project also enables Prevent Storing of IP Addresses server-side so Sentry discards the connection IP address rather than storing it (coarse country/region metadata may still be derived at ingest from the HTTPS connection to sentry.io). See docs/telemetry.md for details.
- Organisation: https://github.com/OpenFrontierLabs
- Manifesto, Architecture, Governance, Credits, Code of Conduct
- Discord, Discussions, Roadmap, Security policy, Glossary, Telemetry
- Cite this work: CITATION.cff
- Founding steward: @bradlishman
Everyone who has helped, in any way, not only code. Add someone by commenting @all-contributors please add @handle for code, docs, ideas on an issue or PR; the bot opens a PR. Compute is credited separately in CREDITS.md.
Error monitoring sponsored by Sentry.
Apache-2.0, see LICENSE. Brand: Frontier Green #66E36E, Compute Green #1B7F46, Graphite #0B0E0D, Silver Gray #6B7177, Soft White #F5F7F6; Inter; dark first.