DeepSeek V4 Pro

DeepSeek V4 Pro is DeepSeek's flagship Mixture-of-Experts model, released open-source under the MIT license on April 24, 2026, with 1.6 trillion total and roughly 49B activated parameters. It uses hybrid attention (Compressed Sparse Attention plus Heavily Compressed Attention) and a reported 1M-token context window, with three reasoning modes: Non-think, Think High and Think Max. Per reporting, its Max configuration scores 80.6% on SWE-bench Verified, the highest open-weights entry, and at 1M context it reportedly needs only about 27% of single-token inference FLOPs and 10% of the KV cache versus V3.2. Ideal for advanced reasoning, complex software engineering and long-running agentic tasks.