DeepSeek-R1
DeepSeek·
LLMsopen-weightFrontier
Overview
First open reasoning model using reinforcement learning. Matched OpenAI o1 on math and coding reasoning tasks without supervised fine-tuning for chain-of-thought.
Capabilities and innovations
671B total / 37B active parameters (MoE)128K context windowChain-of-thought reasoningDistilled variants (1.5B to 70B)Pure RL-based reasoning (no SFT for CoT)Group Relative Policy Optimization (GRPO)Self-verification via reasoning chainsKnowledge distillation to smaller models
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 71.5% | unsourced |
| Humanity’s Last Exam (no tools) | 8.5% | unsourced |
| SWE-bench Verified | 49.2% | unsourced |
| MMLU | 90.8% | unsourced |
Architecture and hardware
- Parameters
- 671B total, 37B active per token (MoE)
- Estimated VRAM at Q4
- ~386 GB, Frontier class
- Quantization formats
- GGUF, GPTQ, AWQ
- Recommended runtime
- Ollama
- License
- MIT
Links
More from DeepSeek
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.