DeepSeek R1: benchmark results
Reasoning model trained to work through problems before answering. Adds supervised cold-start examples before reinforcement learning to address the readability and language-mixing problems of R1-Zero. Provider: DeepSeek. Released 2025-01-20. Access: Open.
Aggregate Rating 1678 ± 3, rank #251 of 1656 rated models, from 1585 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CRMArena - KQA (ReAct) | 61.2 | F1 (%) | 100 |
| CodeGolf Bench - Python Pass@1 | 43.47 | Pass@1 (%): share of problems whose Python solution passes e | 100 |
| CommitSuite - Commit Classification | 63.8 | Accuracy (%, times 100) of ten-way Conventional Commits type | 100 |
| ConceptKT (All Responses) | 14.82 | Missing-concept macro-F1 (%) over the concept labels, histor | 100 |
| ConceptKT (Conceptual Semantic Selection) | 16.87 | Missing-concept macro-F1 (%) over the concept labels, histor | 100 |
| ConceptKT (Same-Concept Only) | 17.4 | Missing-concept macro-F1 (%) over the concept labels, histor | 100 |
| Ducky Bench (Africa) | 1267 | ELO | 100 |
| Enhancing Agent Safety Judgment | 94.53 | Original (Seed) (self-reported) | 100 |
| FinTradeBench (RAG) | 37.7 | Normalized correctness: a calibrated Claude Sonnet 4.5 judge | 100 |
| FinTradeBench (RAG) - Hybrid | 46.4 | Normalized correctness: a calibrated Claude Sonnet 4.5 judge | 100 |
| GaelEval - Gaelic to English Translation (BLEU) | 75.28 | BLEU-4 of Gaelic-to-English translation of An Litir Bheag tr | 100 |
| HELM MedHELM - Patient Communication | 71.88 | Mean win rate | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-r1 · How It Works · Data refreshed daily, snapshot 2026-10-20.