DeepSeek's flagship open-weight Mixture-of-Experts language model with 671B total parameters, 37B active parameters, Multi-head Latent Attention (MLA), and FP8 mixed precision training.
No comparable score for this task lens in the selected independent snapshot.
No comparable score for this task lens in the selected independent snapshot.
No comparable score for this task lens in the selected independent snapshot.
No comparable score for this task lens in the selected independent snapshot.
No comparable score for this task lens in the selected independent snapshot.
Source-backed rows with the original evaluation context preserved.
| Benchmark | Score | Source | Evaluation notes |
|---|---|---|---|
| GPQA / GPQA DiamondChecked 2026-09-11 | 59.1% | DeepSeek AI | GPQA Diamond Pass@1 score from the official DeepSeek-V3 Technical Report. |
| MMLU-ProChecked 2026-09-11 | 75.9% | DeepSeek AI | MMLU-Pro 5-shot accuracy from the official DeepSeek-V3 Technical Report. |
| SWE-bench VerifiedChecked 2026-09-11 | 42% | DeepSeek AI | SWE-bench Verified resolved rate with Agentless framework from the DeepSeek-V3 Technical Report. |
Official DeepSeek-V3 technical report with architecture details (MLA, DeepSeekMoE, FP8) and academic benchmark evaluations.
Official open-weights repository and model card for DeepSeek-V3.
Underlying raw metrics and exact model configurations remain authoritative.
Underlying raw metrics and exact model configurations remain authoritative.