DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948) Jan 20, 2025
Capabilities
- AIME 2024 69.7
- MATH-500 93.9
- GPQA Diamond 59.1
- LiveCodeBench 53.1
- Codeforces rating 1481
DeepSeek-R1-Distill-Qwen-14B
DeepSeek-R1-Distill-Qwen-14B ranks first on MuSR.
31
Benchmarks with a score
1
Ranked first, out of five or more models
1
In the top three
DeepSeek Jan 2025 DeepSeek-R1-Distill-Qwen-14B
Open weights
No Epoch Capabilities Index score yet
- GPQA Diamond 44.7% #177 /229
- BigCodeBench-Hard 20.9% #55 /96
- Open Arabic LLM Leaderboard v1 48% #18 /55
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 | 53.7% | #1 of 98 | Official leaderboard | |
| Open Arabic LLM Leaderboard v1 | 48% | #18 of 55 | Official leaderboard | |
| Seal-0 | 3.60% | #4 of 11 | Official leaderboard | |
| Global PIQA | 70.7% | #22 of 44 | Official leaderboard | |
| MathIF | 39.3% | #7 of 12 | Official leaderboard | |
| ALRAGE | 55.9% | #30 of 39 | Official leaderboard | |
| MadinahQA | 29.1% | #33 of 39 | Official leaderboard | |
| ArbMMLU-HT | 24.1% | #34 of 39 | Official leaderboard | |
| ArabicMMLU | 30.4% | #35 of 39 | Official leaderboard | |
| AlGhafa | 32.6% | #38 of 39 | Official leaderboard | |
| Open Arabic LLM Leaderboard v2 | 32.4% | #38 of 39 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH Level 5 | 87.1% | #21 of 94 | Epoch AI run | |
| PolyMath | 28.6% | #23 of 38 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 50.6% | #136 of 212 | Epoch AI run | |
| MATH-500 | 93.9% | #45 of 70 | Lab model card | |
| AIME 2024 | 69.7% | #66 of 88 | Lab model card | |
| GPQA Diamond | 44.7% | #177 of 229 | Epoch AI run | |
| HMMT 2025 | 31.7% | #45 of 53 | Official leaderboard | |
| AIME 2025 | 49.2% | #46 of 53 | Official leaderboard | |
| BRUMO 2025 | 68.3% | #39 of 42 | Official leaderboard | |
| SMT 2025 | 54.7% | #40 of 40 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench | 53.1% | #26 of 54 | Lab model card | |
| BigCodeBench | 43.3% | #39 of 69 | Official leaderboard | |
| BigCodeBench-Hard | 20.9% | #55 of 96 | Official leaderboard | |
| Codeforces rating | 1,481 | #41 of 47 | Lab model card | |
| Multi-LCB | 16.8% | #11 of 11 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BizFinBench | 59.5% | #10 of 13 | Official leaderboard | |
| BRIDGE | 41.4 | #31 of 39 | Official leaderboard |
Safety and security
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Open LLM Leaderboard v2 | 38.2 | #7 of 99 | 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 | 59.1% GPQA Diamond pass@1, DeepSeek-R1-Distill-Qwen-14B (Table 15) | 44.7% Epoch AI run | +14 pts |
Its system card
Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.