MIT-licensed open-weight model. 1.6T/49B active MoE. 80.6% SWE-bench Verified.
Each score is anchored to human baseline (100). Source URLs link to the original benchmark, leaderboard, or release note.
| Sub-capability | Quality | Source | Score vs. baseline | Score |
|---|---|---|---|---|
| Logical Reasoning | vendor reported | deepseek.com/v4 | 74% | |
| Multi-step Planning | vendor reported | deepseek.com/v4 | 58% |