DeepSeek-V2
DeepSeek·
LLMsopen-weightcloud + localMulti-GPU Self-Host
Overview
Strong performance at a lower cost.
Capabilities and innovations
MLA ArchitectureGPT-4 classExtremely cheap APIMulti-head Latent Attention (MLA)KV Cache Compression (93.3%)DeepSeekMoE Architecture
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 48.3% | unsourced |
| Humanity’s Last Exam (no tools) | 6.1% | unsourced |
| SWE-bench Verified | 18.9% | unsourced |
| MMLU | 77.8% | unsourced |
- SWE-bench Pro: benchmark did not exist at release
- MMLU-Pro: benchmark did not exist at release
- Terminal-Bench 2.x: benchmark did not exist at release
- MMMU-Pro: benchmark did not exist at release
Architecture and hardware
- Parameters
- 236B total, 21B active per token (MoE)
- Estimated VRAM at Q4
- ~136 GB, Multi-GPU Self-Host class
Links
More from DeepSeek
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.