DeepSeek-V4-Pro-Max

DeepSeek-V4-Pro-Max is the maximum reasoning effort mode of DeepSeek-V4-Pro, a 1.6T-parameter MoE model with 49B activated parameters and a 1M-token context window. It introduces a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for dramatically improved long-context efficiency, requiring only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2 at 1M-token context. The model also incorporates Manifold-Constrained Hyper-Connections (mHC) for stable signal propagation and is trained with the Muon optimizer for faster convergence. Pre-trained on more than 32T tokens, V4-Pro-Max significantly advances open-source knowledge capabilities, achieves top-tier performance in coding benchmarks, and bridges the gap with leading closed-source models on reasoning and agentic tasks.