DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948) Jan 20, 2025
Capabilities
- AIME 2024 72.6
- MATH-500 94.3
- GPQA Diamond 62.1
- LiveCodeBench 57.2
- Codeforces rating 1691
DeepSeek-R1-Distill-Qwen-32B
DeepSeek-R1-Distill-Qwen-32B places in the top three on BigO(Bench).
39
Benchmarks with a score
0
Ranked first, out of five or more models
1
In the top three
DeepSeek Jan 2025 DeepSeek-R1-Distill-Qwen-32B
Open weights
No Epoch Capabilities Index score yet
- GPQA Diamond 64.1% #131 /229
- OCR-Reasoning 26.5% #8 /15
- AgentIF 55.1% #9 /12
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MuSR | 45.3% | #14 of 98 | Official leaderboard | |
| MathIF | 42.6% | #4 of 12 | Official leaderboard | |
| Global PIQA | 72.2% | #19 of 44 | Official leaderboard | |
| MT-AIME2024 | 31% | #2 of 4 | Official leaderboard | |
| BBEH micro average | 19.2% | #6 of 11 | Official leaderboard | |
| BIG-Bench Extra Hard | 5.20% | #6 of 11 | Official leaderboard | |
| ALRAGE | 61.7% | #26 of 39 | Official leaderboard | |
| AlGhafa | 46.5% | #29 of 39 | Official leaderboard | |
| AgentIF | 55.1% | #9 of 12 | Official leaderboard | |
| MadinahQA | 29.4% | #31 of 39 | Official leaderboard | |
| Open Arabic LLM Leaderboard v2 | 35.1% | #34 of 39 | Official leaderboard | |
| ArabicMMLU | 29.8% | #37 of 39 | Official leaderboard | |
| ArbMMLU-HT | 23.4% | #39 of 39 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| PolyMath | 35.1% | #17 of 38 | Official leaderboard | |
| GPQA Diamond | 64.1% | #131 of 229 | Epoch AI run | |
| MATH-500 | 94.3% | #43 of 70 | Lab model card | |
| OTIS Mock AIME 2024-2025 | 55.6% | #131 of 212 | Epoch AI run | |
| AIME 2024 | 72.6% | #57 of 88 | Lab model card | |
| AIME 2025 | 60% | #42 of 53 | Official leaderboard | |
| HMMT 2025 | 33.3% | #44 of 53 | Official leaderboard | |
| PHYSICS | 6.80% | #18 of 20 | Official leaderboard | |
| PHYBench | 11.7% | #26 of 28 | Official leaderboard | |
| SMT 2025 | 60.4% | #38 of 40 | Official leaderboard | |
| BRUMO 2025 | 68.3% | #40 of 42 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BigO(Bench) | 41.1% | #2 of 18 | Official leaderboard | |
| BigCodeBench | 49.4% | #24 of 69 | Official leaderboard | |
| BigCodeBench-Hard | 26.4% | #38 of 96 | Official leaderboard | |
| LiveCodeBench | 57.2% | #24 of 54 | Lab model card | |
| Codeforces rating | 1,691 | #37 of 47 | Lab model card | |
| Multi-LCB | 22% | #9 of 11 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BALROG | 19.5% | #21 of 33 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BizFinBench | 66.3% | #6 of 13 | Official leaderboard | |
| BRIDGE | 44.3 | #25 of 39 | Official leaderboard | |
| BizFinBench v2 | 31.3% | #8 of 9 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| OCR-Reasoning | 26.5% | #8 of 15 | Official leaderboard |
Safety and security
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Open LLM Leaderboard v2 | 23.0 | #47 of 99 | Official leaderboard | |
| Political Science LLM Benchmark | 63.4% | #37 of 42 | Official leaderboard |
Its lab's number next to a leaderboard
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| GPQA Diamond | 62.1% GPQA Diamond pass@1, DeepSeek-R1-Distill-Qwen-32B (Table 15) | 64.1% Epoch AI run | −2.0 pts |
Its system card
Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.