DeepSeek-R1
More info
- Creator
- DeepSeek AI
- Released
- Owner
- DeepSeek
- Type
- Reasoning LLM
- License
- MIT (weights)
- Website
- deepseek.com
- Source
- github.com/deepseek-ai/DeepSeek-R1
- Wikipedia
- en.wikipedia.org
Chinese open-weights reasoning model (DeepSeek, January 2025) that matched OpenAI o1 on maths and coding — for about one-hundredth of the training cost.
Proved that pure RL (no SFT warm-up) can produce strong reasoning. The predecessor R1-Zero was trained purely on reward signals for correct maths/code answers; the final R1 was distilled into smaller models (Qwen-7B, Llama-8B) that suddenly punched far above their weight.
Market shock: shattered the assumption that frontier models required billions and closed data. Nvidia's stock fell ~17 % in a day. Weights released under MIT licence on Hugging Face.