Open-R1: a fully open reproduction of DeepSeek-R1
Hugging Face
Hugging Face is launching Open-R1 to rebuild DeepSeek-R1 from scratch, in the open. The model shocked everyone last week, but its training data and code stayed secret.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek-R1 landed like a grenade in the AI world last week. It matched or beat OpenAI's o1 on math, coding, and logic tasks, and it came with a tech report explaining, in surprising detail, how the team pulled it off. The core trick was reinforcement learning applied to a base model with essentially no human hand-holding on reasoning steps. That's a big deal because until now the actual recipes behind reasoning models like o1 were locked away as trade secrets.
But a report is not a codebase, and DeepSeek released weights without the underlying data or training scripts. That's the gap Hugging Face wants to close with Open-R1, a project aimed at reverse-engineering the whole pipeline so anyone can reproduce it, tweak it, or improve on it.
The DeepSeek approach itself is worth understanding. DeepSeek-V3, the 671-billion-parameter Mixture of Experts base model, reportedly cost just $5.5 million to train, thanks to tricks like Multi Token Prediction and Multi-Head Latent Attention. From there, DeepSeek built two variants. R1-Zero skipped supervised fine-tuning entirely and learned to reason purely through reinforcement learning using a method called Group Relative Policy Optimization, guided by simple accuracy-and-structure rewards. It worked, but the outputs were often messy and hard to read. R1 fixed that by adding a small
My take — AI-written commentary, not fact-checked reporting
,
Read more about this at: Hugging Face