DeepSeek Launches Open-Source V4 Pro and Flash with Million-Token Context
DeepSeek released two models: V4-Pro with 1.6T parameters (49B activated) and V4-Flash with 284B parameters (13B activated), both featuring a hybrid attention architecture that reduces compute per token to 27% of the previous generation. In maximum thinking mode, Pro-Max surpassed GPT-5.4 with a Codeforces Elo of 3206 and topped LiveCodeBench at 93.5%. Both models are available as open-source previews on Hugging Face.