DeepSeek-V4-Pro-0813
DeepSeek·
LLMsopen-weightFrontier
Overview
General-availability build of DeepSeek-V4-Pro. Identical 1.6T total / 49B active MoE architecture and size as the April preview — re-post-trained, with the gains concentrated in agentic and software-engineering tasks. 1M-token context, MIT license.
Capabilities and innovations
1.6T total / 49B active MoE1M token context windowThree reasoning effort modes (incl. Think Max)MIT licenseHybrid attention: CSA + HCAManifold-Constrained Hyper-ConnectionsRe-post-training of the April preview checkpoint
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 92.83% | third party GPQA Diamond 92.83%, Artificial Analysis, reasoning max effort, measured on the 0813 build (April preview: 89.4 on vals.ai, 90.1 vendor). Same source lineage as the V4-Flash-0731 entry. |
| Humanity’s Last Exam (no tools) | 39.34% | third party HLE 39.34% no-tools, Artificial Analysis Intelligence Index component, max effort (April preview: 37.7 vendor). |
| SWE-bench Verified | 96.4% | independent SWE-bench Verified 96.40 (± 0.83), vals.ai, max effort (April preview: 77.4 on the same harness). Era-1 benchmark, near saturation — recorded for continuity, not ranked. |
| Terminal-Bench 2.x | 54.68% | independent Terminal-Bench 2.1 54.68%, measured by Vals AI with the Terminus 2 harness (accessed 2026-09-23). Vals AI is the reference harness for this field so values compare across vendors; vendor and other third-party figures stay in the notes below. Replaces 78.65 (third_party). |
| ARC-AGI-1 | 90.5% | independent ARC-AGI-1 90.5% Semi-Private, verified by ARC Prize (Low effort, the highest of four variants; Max: 90.0), tested 2026-08-13 (accessed 2026-09-23). |
| ARC-AGI-2 | 61.3% | independent ARC-AGI-2 61.3% Semi-Private, verified by ARC Prize (Max effort, $0.60/task; High: 59.7, Low: 56.3), tested 2026-08-13 (accessed 2026-09-23). |
- SWE-bench Pro: independent evaluation pending — The April preview checkpoint scored 55.4; no re-run published for the 0813 build.
- MMLU-Pro: independent evaluation pending — The April preview checkpoint scored 87.5 (vendor) / 87.25 (vals.ai); vals.ai has not yet re-run MMLU Pro on the 0813 build.
- MMMU-Pro: not reported by the vendor
Architecture and hardware
- Parameters
- 1600B total, 49B active per token (MoE)
- Estimated VRAM at Q4
- ~920 GB, Frontier 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.