• Hugging Face has launched Open-R1 to replicate DeepSeek’s R1 AI model and fully open-source its training process, including datasets and methods.
  • DeepSeek’s R1 model fact-checks itself, making it more reliable in fields like science and math. Its chatbot app quickly topped the Apple App Store charts.

Barely a week after DeepSeek’s R1 “reasoning” AI model disrupted the market, researchers at Hugging Face are launching Open-R1—an initiative to replicate the model from scratch and fully open-source its components.

Hugging Face’s head of research, Leandro von Werra, and engineers aim to counter DeepSeek’s “black box” approach by publicizing all datasets, experiment details, and training methods. While DeepSeek’s R1 is technically “open” with a permissive license, key aspects of its training process remain undisclosed, making it difficult for researchers to build upon.

“The R1 model is impressive, but there’s no open dataset, experiment details, or intermediate models available, which makes replication and further research difficult,” said Hugging Face engineer Elie Bakouch.
“Fully open sourcing R1’s complete architecture isn’t just about transparency — it’s about unlocking its potential.”

DeepSeek, a Chinese AI lab backed by a quantitative hedge fund, released R1 last week, quickly surpassing OpenAI’s o1 model on multiple reasoning benchmarks. Unlike typical AI models, R1 fact-checks itself, making it more reliable for complex problem-solving in science, math, and physics. DeepSeek’s chatbot app, powered by R1, shot to the top of Apple’s App Store charts, sparking debates over U.S. competitiveness in AI.

Open-R1 is less concerned with geopolitical dominance and more focused on advancing AI transparency.

“Having control over the dataset and process is critical for deploying a model responsibly in sensitive areas,” Bakouch noted.

Hugging Face’s Science Cluster, equipped with 768 Nvidia H100 GPUs, will drive Open-R1’s replication efforts. The project has already gained traction, amassing 10,000 GitHub stars in three days.

“When the R1 recipe has been replicated, anyone who can rent some GPUs can build their own variant of R1 with their own data,” said Bakouch. “It’s an important shift that strengthens openness in AI.”

Edited by Harshajit Sarmah