DeepSeek-V3.2
DeepSeek·
LLMsopen-weightcloud + localFrontier
Overview
Successor to DeepSeek V3. 671B MoE (37B active) with DeepSeek Sparse Attention for long-context efficiency and large-scale agentic tool-use pipeline.
Capabilities and innovations
671B total / 37B active MoE164K contextDeepSeek Sparse Attention (DSA)Agentic tool-use (1,800+ environments)Strong reasoning without thinking modeMIT licenseDeepSeek Sparse Attention (DSA)Large-Scale Task Synthesis Pipeline1,800+ Agentic Environments
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| MMLU-Pro | 84.92% | unsourced |
- GPQA Diamond: not reported by the vendor
- Humanity’s Last Exam (no tools): not reported by the vendor
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: not reported by the vendor
- MMLU: not reported by the vendor
- Terminal-Bench 2.x: benchmark did not exist at release
- MMMU-Pro: benchmark did not exist at release
Architecture and hardware
- Parameters
- 671B total, 37B active per token (MoE)
- Estimated VRAM at Q4
- ~386 GB, Frontier class
Reliability
- Hallucination rate (Vectara HHEM)
- 6.3% (lower is better)independent
- Agentic tool use (τ-bench)
- 71.3% (higher is better)third party
Links
More from DeepSeek
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.