Open reasoning model trained with large-scale reinforcement learning, with distilled variants.
Specifications
- Provider
- DeepSeek
- Type
- Open-source / open-weight
- Modality
- Reasoning
- Category
- Reasoning model
- Context window
- 128K
- License
- MIT
- Knowledge cutoff
- 2024
- Released
- January 20, 2025
What it was trained for
Step-by-step reasoning across math, code, and logic, trained with large-scale reinforcement learning.
Best for
- ▸Complex math problem solving
- ▸Competitive and algorithmic coding
- ▸Multi-step logical reasoning
- ▸Research and analysis tasks
Capabilities
Long chain-of-thought reasoning with visible tracesSelf-verification during reasoningOpen weights (MIT)Distilled smaller variants available
Performance & positioning
A leading open reasoning model whose chain-of-thought quality is competitive with top proprietary reasoning models, at the cost of longer inference time.
More from DeepSeek