Research Vault

DeepSeek-V4-Pro-0813

Properties1
tagshuggingface, text-generation, coding, agentic, moe, deepseek-v4, mla, fp8, million-token-context, reasoning, instruct, dspark

3 min read

DeepSeek-V4-Pro-0813

One-line: GA of DeepSeek-V4-Pro (supersedes the April preview) with a major agentic upgrade — Terminal Bench 2.1 87.9 (+15.8pp vs preview), DeepSWE 62.7 (from 12.8), Toolathlon 74.1 — shipped under MIT with official FP8 weights and a built-in DSpark speculative-decoding module. ⚠ large: 1.6T total params, server-class only. ⚠ no SWE-bench Verified / BFCL published — agentic numbers come from the lab’s DeepSeek-Harness suite (lab-reported).

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published for 0813
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published for 0813
Aider polyglot—not published
SWE-bench Verified—not published (agentic stand-in: DeepSWE below)
SWE-bench Multilingual—not published
Terminal Bench 2.187.9lab-reported (preview 72.1, Flash-0731 82.7, GLM-5.2 81.0, Kimi K3 88.3, Opus-4.8 85.0)
NL2Repo61.5lab-reported (preview 38.5)
Cybergym83.3lab-reported (preview 52.7, Flash-0731 76.7)
DeepSWE62.7lab-reported (preview 12.8 — biggest single jump)
DSBench-FullStack †71.1lab-reported, internal test set (preview 41.8)
DSBench-Hard †67.2lab-reported, internal test set (preview 31.1)

† Internal full-stack / hard coding-agent test sets — treat as lab-internal, not leaderboard-comparable.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL—not published
τ-Bench—not published
ToolACE—not published
GAIA—not published
Toolathlon-Verified74.1lab-reported (preview 55.9, Flash-0731 70.3)
Agents’ Last Exam25.7lab-reported (preview 16.5)
AutomationBench (Public)31.8lab-reported (preview 12.8)

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published for 0813
GPQA Diamond—not published for 0813
MATH—not published
AIME 2025—not published
HLE (wo / w tools)42.7 / 60.0lab-reported (preview 37.7 / 48.2)

Efficiency

  • Active params: ~49B per token (top-6 of 384 routed + 1 shared) — same structure as V4-Pro preview (config verified identical: 61 layers, hidden 7168)
  • Total params: ~1.6T (full disk footprint)
  • Context length: 1,048,576 tokens (1M) via YaRN ×16 from 64k native; recommends ≥384K output budget for high/max reasoning effort
  • VRAM (fp16 rough): ~98 GB active-params equivalent at runtime; full weights ~1.6 TB on disk — official weights are FP8 e4m3 (quantization_config in config.json), no BF16 variant published yet
  • Architecture: custom deepseek_v4 (DeepseekV4ForCausalLM) — needs trust-remote-code; hybrid CSA+HCA attention, hash layers + nextn-predict layer, DSpark speculative decoding module (--speculative-config '{"method":"dspark",...}' in vLLM / --speculative-algorithm DSPARK in SGLang)
  • Deployment: vLLM recipe targets a 4×GB300 node; local single-machine path exists via the inference/ folder (weight conversion + chat demo)

What makes it notable

This is the production GA of DeepSeek-V4-Pro, and the delta over the April preview is the story: +15.8pp on Terminal Bench 2.1 (87.9 vs 72.1), DeepSWE 12.8 → 62.7, Cybergym 52.7 → 83.3, Toolathlon 55.9 → 74.1 — the lab’s own table now puts it broadly competitive with Kimi K3 (Terminal Bench 88.3, DeepSWE 67.5) and GLM-5.2 (81.0 / 46.2), within a few points of Opus-4.8 and Fable-5 (w/ fallback) on the same suite. Like the preview it ships MIT with 1M-token context, and the DSpark speculative module is now bundled in the same checkpoint (no separate draft model). Caveats for agent-loop users: no SWE-bench Verified or BFCL published — the agentic numbers are all lab-reported via DeepSeek Harness, two of them (DSBench) on internal test sets; and an early HN report flags pass@1 reliability concerns (works well at pass@3, “horrendous” at pass@1 with temperature 1.0) — sample or lower temperature in production. It is emphatically not a local model: 1.6T total params, FP8 weights, multi-node deployment. Replaces the API preview build; OpenRouter lists it as the GA.

See also

Graph View

Backlinks