Hugging Face launches Open-R1 to reproduce DeepSeek-R1

Hugging Face launched Open-R1, a project to reconstruct the parts of DeepSeek-R1’s training pipeline that DeepSeek itself had not released. DeepSeek-R1, published a week earlier, had shipped as downloadable weights under an MIT licence with a paper describing its method, but — as with almost every open-weight release before it — the underlying training data, code and hyperparameters stayed private. Open-R1 set out to fill that gap: reasoning datasets for maths, code and logic, the training code itself, and an account of the scaling laws and compute-to-data trade-offs behind reasoning models, released openly rather than reverse-engineered from the paper alone.

The project described a three-stage plan mirroring DeepSeek’s own reported pipeline. First, distil high-quality reasoning traces from DeepSeek-R1 to produce a set of R1-Distill models trained by supervised fine-tuning. Second, reproduce R1-Zero, the more radical of DeepSeek’s two models, trained by reinforcement learning against verifiable answers with no supervised stage at all. Third, replicate the full multi-stage pipeline DeepSeek described — base model, supervised fine-tuning, then reinforcement learning — to test whether the published paper’s account of the method actually reproduced its results.

Hugging Face framed the announcement as a starting point rather than a finished reproduction, and invited outside contributors to help build the missing datasets and code. Open-R1’s premise — that a release accompanied only by weights and a paper leaves the field unable to verify or build on the underlying method — became one of the more concrete tests of DeepSeek’s contribution: within weeks, other groups were publishing partial reproductions and variations of R1’s reinforcement-learning recipe built on Open-R1’s early releases.