DeepSeek R1
Last updated: August 10, 2026
Specs
| Vendor | DeepSeek |
|---|
| Released | January 20, 2025 |
|---|
| Context window | 128K tokens |
|---|
| Input | $0.7 / M tokens |
|---|
| Output | $2.5 / M tokens |
|---|
Benchmarks
| mmlu | 90.8 |
|---|
| gpqa | 71.5 |
|---|
| humanEval | 90.2 |
|---|
| arenaElo | 1389 |
|---|
| sweBench | 49.2 |
|---|
| sweBenchPro | 38 |
|---|
| liveCodeBench | 65 |
|---|
| aiderPolyglot | 57 |
|---|
| terminalBench | 30 |
|---|
| mmluPro | 84 |
|---|
| hle | 9.4 |
|---|
| aime | 79.8 |
|---|
| math500 | 97.3 |
|---|
Strengths
- Frontier reasoning at open-weight price
- Exceptional math and STEM benchmarks
- Can be self-hosted
Weaknesses
- No native multimodal support
- Tool use less polished than OpenAI/Anthropic
Best for
- Math, science and engineering problem solving
- Cost-sensitive reasoning workloads
- Research teams needing local inference
Compare with other AI models