DeepSeek-V4-Pro: Open-Source Coding SOTA, 1.6T Params, 1M Context, MIT License
DeepSeek released the V4 series on April 24 — two models: V4-Pro (1.6T total params, 49B activated) and V4-Flash (284B total, 13B activated). Both support a 1 million token context window. Weights are MIT licensed and available immediately on HuggingFace.
Benchmark Numbers
V4-Pro-Max (maximum reasoning mode) vs frontier models:
| Eval | Opus 4.6 Max | GPT-5.4 xHigh | Gemini 3.1 Pro High | DS-V4-Pro Max |
|---|---|---|---|---|
| LiveCodeBench | 88.8% | — | 91.7% | 93.5% |
| Codeforces Rating | — | 3168 | 3052 | 3206 |
| SimpleQA-Verified | 46.2% | 45.3% | 75.6% | 57.9% |
| GPQA Diamond | 91.3% | 93.0% | 94.3% | 90.1% |
| HLE | 40.0% | 39.8% | 44.4% | 37.7% |
| SWE-Verified | 80.8% | — | 80.6% | 80.6% |
| SWE-Pro | 57.3% | 57.7% | 54.2% | 55.4% |
| Terminal-Bench 2.0 | 65.4% | 75.1% | 68.5% | 67.9% |
| GDPval-AA (ELO) | 1619 | 1674 | 1314 | 1554 |
LiveCodeBench 93.5% is the strongest reported result at the frontier on that benchmark. Codeforces 3206 similarly leads the field. Both are pure coding signal, not self-reported.
Where V4-Pro-Max trails: Terminal-Bench (10 points behind GPT-5.5 and 7 behind GPT-5.4), GPQA Diamond, and HLE — the hardest scientific reasoning evals where Gemini 3.1 Pro currently leads.
Architecture Changes
Three key upgrades over V3.2:
Hybrid Attention (CSA + HCA): At 1M-token context, V4-Pro requires only 27% of V3.2’s per-token inference FLOPs and 10% of its KV cache. This is a substantial efficiency gain for long-context deployment, not a marginal one.
Manifold-Constrained Hyper-Connections (mHC): Replaces standard residual connections. DeepSeek’s claim is improved signal propagation without sacrificing expressivity.
Muon Optimizer: Used in training for faster convergence and stability. Muon has been gaining traction in frontier labs after showing strong results in smaller-scale ablations.
Pre-training used more than 32T tokens. Post-training runs in two stages: domain-specific SFT and RL with GRPO, followed by on-policy distillation to merge the specialised experts back into one model.
Three Reasoning Modes
Both V4-Pro and V4-Flash expose three inference modes:
- Non-Think: Fast, no chain-of-thought
- Think High: Standard reasoning budget
- Think Max: Extended reasoning — requires at least 384K context window
The benchmark table above uses Think Max. Non-Think mode scores drop substantially (LiveCodeBench falls from 93.5% to 56.8% for Flash; V4-Pro Non-Think hasn’t been benchmarked publicly yet on coding evals).
Open Source and Pricing
Weights are MIT licensed. The model uses FP4 + FP8 mixed precision — MoE expert params in FP4, most other weights in FP8. API pricing has not been announced. DeepSeek historically prices aggressively relative to closed labs; V3.2 is currently available via the DeepSeek API at a fraction of GPT-5 pricing.
What This Means
V4-Pro is the strongest open-source coding model available today by the benchmarks that matter most. For teams running self-hosted inference, the 27% FLOP reduction at 1M context is operationally significant — it changes the unit economics of long-context workloads substantially.
The gap to GPT-5.5 on Terminal-Bench (82.7% vs 67.9%) is real and not trivial. But V4-Pro is competitive on SWE evals, dominant on LiveCodeBench, and available to run locally without an API relationship. That combination hasn’t existed before at this capability level.