DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (arXiv 2501.12948) Jan 20, 2025
Capabilities
- AIME 2024 70
- MATH-500 94.5
- GPQA Diamond 65.2
- LiveCodeBench 57.5
- Codeforces rating 1633
DeepSeek-R1-Distill-Llama-70B
DeepSeek-R1-Distill-Llama-70B ranks first on BigO(Bench).
28
Benchmarks with a score
1
Ranked first, out of five or more models
1
In the top three
DeepSeek Jan 2025 DeepSeek-R1-Distill-Llama-70B
Open weights
No Epoch Capabilities Index score yet
- GPQA Diamond 55.7% #146 /229
- AgentIF 55% #10 /12
- BigCodeBench-Hard 20.3% #59 /96
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 | |
|---|---|---|---|---|
| Open Arabic LLM Leaderboard v1 | 53.6% | #8 of 55 | Official leaderboard | |
| MuSR | 43.4% | #30 of 98 | Official leaderboard | |
| MathIF | 41.4% | #5 of 12 | Official leaderboard | |
| AgentIF | 55% | #10 of 12 | Official leaderboard | |
| AbstentionBench | 58.2% | #5 of 5 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH Level 5 | 89.9% | #17 of 94 | Epoch AI run | |
| 3SAT-Search | 72.2% | #4 of 11 | Official leaderboard | |
| TSP-Opt | 32.4% | #5 of 11 | Official leaderboard | |
| PolyMath | 29.4% | #21 of 38 | Official leaderboard | |
| Misguided Attention | 42.6% | #26 of 45 | Official leaderboard | |
| MATH-500 | 94.5% | #42 of 70 | Lab model card | |
| OTIS Mock AIME 2024-2025 | 51.4% | #135 of 212 | Epoch AI run | |
| GPQA Diamond | 55.7% | #146 of 229 | Epoch AI run | |
| AIME 2024 | 70% | #64 of 88 | Lab model card | |
| AIME 2025 | 55% | #43 of 53 | Official leaderboard | |
| HMMT 2025 | 33.3% | #43 of 53 | Official leaderboard | |
| SMT 2025 | 60.9% | #37 of 40 | Official leaderboard | |
| BRUMO 2025 | 66.7% | #41 of 42 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BigO(Bench) | 41.4% | #1 of 18 | Official leaderboard | |
| LiveCodeBench | 57.5% | #23 of 54 | Lab model card | |
| BigCodeBench | 42.6% | #41 of 69 | Official leaderboard | |
| BigCodeBench-Hard | 20.3% | #59 of 96 | Official leaderboard | |
| Codeforces rating | 1,633 | #40 of 47 | Lab model card | |
| SWE-PolyBench | 5.31% | #6 of 6 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SnakeBench | 30 | #37 of 188 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BRIDGE | 46.2 | #21 of 39 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| BaxBench | 5.94% | #29 of 34 | Official leaderboard |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Open LLM Leaderboard v2 | 27.8 | #30 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 | 65.2% GPQA Diamond pass@1, DeepSeek-R1-Distill-Llama-70B (Table 15) | 55.7% Epoch AI run | +9.5 pts |
Its system card
Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.