DeepSeek R1-Zero
R1-Zero studies whether reasoning behavior can emerge from reinforcement learning alone. It uses the full DeepSeek-V3-Base foundation and has published weights.
Research release
Correctness and output format supply the training signal
R1-Zero uses Group Relative Policy Optimization, which compares scores within groups of outputs instead of training a separate critic. Rule-based rewards check answer correctness and reasoning-tag format. The researchers avoided a learned neural reward model because of reward hacking and added training cost.DeepSeek-R1 research paper and author record
Reasoning behavior did not solve readability
The release reports self-checking and reflection, but also repetition, poor readability and mixed languages. R1 added supervised cold-start examples to address these weaknesses.DeepSeek-R1-Zero model card
Zero retains the large DeepSeek architecture
The published configuration uses DeepseekV3ForCausalLM with 61 layers and eight routed experts per token. Its 163,840 max_position_embeddings setting is a file-level value, not a separate advertised context specification.R1-Zero checkpoint configuration
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | DeepSeek-R1-Zerodeepseek-ai/DeepSeek-R1-Zero |
| Context window | 131,072 tokens Published 128K context for the weight model. DeepSeek-R1-Zero model card |
| Parameters | 671 billion total; 37 billion active per token.DeepSeek-R1-Zero model card |
| Weight license | MIT.DeepSeek announcement of R1 and R1-Zero, 20 January 2025 |
Related models
- DeepSeek R1-0528
DeepSeek · DeepSeek R1
Updated full-size R1 reasoning model with more post-training computation, deeper reasoning and revised prompt handling.
Launch date Downloadable - DeepSeek R1
DeepSeek · DeepSeek R1
Reasoning language model that combines supervised examples and reinforcement learning on a large mixture-of-experts foundation.
Launch date Downloadable