DeepSeek-V4-Pro notes: flagship 1M-context agent model

DeepSeek-V4-Pro is the flagship I treat as the Expert-Mode brain. The April 2026 preview and the later “official version” banner agree on the shape: 1.6 trillion total parameters, 49 billion active, 1M context, Thinking and non-Thinking, Responses API, and Codex integration. The launch note claims open-source state of the art on agentic coding benches, strong world knowledge, and math/STEM/coding that they put next to closed flagships. I did not rerun those benches. I did run it on a messy service and it earned the Pro name.

On the API the model string is deepseek-v4-pro. The docs I opened mapped that name to DeepSeek-V4-Pro-0813. Context length 1M, max output 384K, JSON, tool calls, Responses API, Anthropic API, FIM in non-thinking only. Concurrency on the pricing table was 500. Peak hours cost more; I still check the live table before I quote a unit price in a budget meeting.

Chat exposes this path as Expert Mode. If you want the lighter sibling, read DeepSeek-V4-Flash. For picture input, that is a different experimental name: DeepSeek-V4-Flash-Vision-Exp.