DeepSeek-V4-Pro-0813 is the official GA build of DeepSeek's flagship model, released August 12, 2026. A Mixture-of-Experts transformer with 1.6 trillion total parameters (49 billion active), it reached general availability through a quiet, build-tagged release on the DeepSeek API. The 0813 build succeeded the April preview the same day. Weights are MIT-licensed on Hugging Face.

Model Specifications

Specification	DeepSeek V4 Pro 0813
Architecture	DeepSeekMoE (Mixture-of-Experts)
Total parameters	1.6T
Active parameters	49B
Context window	1,000,000 tokens (max output 384K)
Pre-training data	33T tokens
Precision	FP4 (routed experts) + FP8 mixed
Modality	Text only
Reasoning modes	Non-Think / Think High / Think Max (switchable per request)
License	MIT (open weights)

Key Features

World-leading coding: LiveCodeBench 93.5% (world #1) and Codeforces rating 3206 - roughly the level of the top-23 humans.
Efficient 1M context: Hybrid attention (CSA + HCA) uses only 27% of single-token inference FLOPs and 10% of the KV cache vs DeepSeek-V3.2 at 1M context.
Architectural innovations: Manifold-Constrained Hyper-Connections (mHC), Muon optimizer, and Multi-Token Prediction.
Unified reasoning modes: Non-Think, Think High, and Think Max are selectable per request within a single API model.
Broad API compatibility: OpenAI ChatCompletions, Anthropic Messages, and Responses API.

Benchmark Highlights

SWE-bench Verified: 80.6% (+12.8 pts vs V3.2)
LiveCodeBench: 93.5% (world #1)
Codeforces: 3206 Elo
GPQA Diamond: 90.1%
MMLU-Pro: 87.5%
Terminal-Bench 2.1: 87.9 (up from 72.1 in the April preview)
HMMT 2026 Feb: 95.2%
LMArena: 1456 Elo (June 2026)

Pricing

Tier	Cache Hit (input)	Cache Miss (input)	Output
deepseek-v4-pro	$0.003625	$0.435	$0.87

With a blended rate around $0.65 per million tokens, DeepSeek V4 Pro is roughly 46× cheaper than Claude Fable 5 ($10/$50 per million tokens) while trailing it by only ~5.3% on average across nine comparable agent benchmarks. DeepSeek has posted notice that prices will increase in the near future.

Notes

The 0813 build shipped without a blog post or fanfare; model name on the pricing page simply changed to the -0813 build tag. Independent replication of the 0813-specific benchmarks is pending - the April preview weights were still the ones hosted on Hugging Face as of August 13.

Conclusion

DeepSeek V4 Pro 0813 delivers frontier-level coding and agentic performance in an MIT-licensed, 1.6T MoE package at a fraction of closed-model prices. For developers who prioritize value, open weights, and long-context agent work, it remains one of the most compelling models on the market.

Related: Claude 3.5 Sonnet and Claude 3 Haiku. Hub: models.