DeepSeek-V4-Flash-0731
DeepSeek·
LLMsopen-weightMulti-GPU Self-Host
Overview
General-availability build of DeepSeek-V4-Flash. Identical 284B total / 13B active MoE architecture and size as the April preview — re-post-trained, with the gains concentrated in agentic tool use. 1M-token context, MIT license. On 2026-08-21 DeepSeek added a vision variant on top of this checkpoint, DeepSeek-V4-Flash-Vision-Exp — API-only, without open weights.
Capabilities and innovations
284B total / 13B active MoE1M token context windowThree reasoning effort modes (incl. Think Max)Agentic tool use (Terminal-Bench, NL2Repo, Cybergym, Toolathlon)MIT licenseHybrid attention: CSA + HCAManifold-Constrained Hyper-ConnectionsAgent-focused re-post-training on the preview checkpoint
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 91% | independent GPQA Diamond 91 — Artificial Analysis, measured on the 0731 build (+1 point over the April V4 Flash). |
| Humanity’s Last Exam (no tools) | 37% | independent HLE 37 no-tools — Artificial Analysis Intelligence Index component (+5 points over the April V4 Flash). |
| Terminal-Bench 2.x | 79% | independent Terminal-Bench 2.1 79.0 — Artificial Analysis (+17 points over the April V4 Flash). DeepSeek self-reports 82.7 on the same benchmark version; the independent value is used as primary. Version 2.1, not comparable to the 2.0 scores recorded for earlier models. |
- SWE-bench Verified: independent evaluation pending — The 0731 changelog reports only agentic suites. The preview checkpoint's 79.0 SWE-bench Verified is not carried over.
- SWE-bench Pro: independent evaluation pending — Preview checkpoint scored 52.6; no re-run published for 0731.
- MMLU-Pro: independent evaluation pending — Preview checkpoint scored 86.2; no re-run published for 0731.
- MMMU-Pro: not reported by the vendor
Architecture and hardware
- Parameters
- 284B total, 13B active per token (MoE)
- Estimated VRAM at Q4
- ~164 GB, Multi-GPU Self-Host class
- Quantization formats
- FP8, GGUF
- Recommended runtime
- vLLM
- License
- MIT
Links
More from DeepSeek
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.