DeepSeek V4-Pro
DeepSeekTEXT Open weight
- Provider
- DeepSeek
- Modality
- TEXT
- Parameters
- 1600B
- Context window
- 1,000,000 tokens
- Weights
- Open
- Released
- 24 Apr 2026
DeepSeek V4-Pro (April 2026, preview) is DeepSeek's MIT-licensed frontier flagship — a ~1.6T/49B-active MoE with a 1M context and built-in reasoning that subsumes the R-series.
Best for
- Frontier coding / agentic work (tops the line on LiveCodeBench, SWE-bench)
- Hardest reasoning/science via built-in thinking mode
- True 1M-context tasks (whole-repo, long-doc)
- When quality matters more than cost
How it compares — and the India angle
- MIT-licensed frontier weights give data residency for Indian firms, but 1.6T MoE means most users access it via API, not self-host.
- Reasoning is folded into V4's hybrid thinking mode — there is no standalone R2.
- Preview status: full benchmark tables aren't yet in primary sources — treat blog claims as approximate.
Benchmarks
Representative public scores (approximate, higher is better) for relative comparison. Check the provider for the latest official results.
GPQA Diamond90
SWE-bench Verified81
How to access
API ~$0.44 / 1M cache-miss input (~$0.004 cache-hit), ~$0.87 / 1M output. MIT (free to self-host; large cluster).
Access DeepSeek V4-Pro