The DeepSeek-V4 series introduces next-generation Mixture-of-Experts (MoE) open language models engineered for advanced reasoning and massive long-context efficiency. Natively supporting a 1-million token context window, the lineup includes the flagship DeepSeek-V4-Pro (1.6T total / 49B activated parameters) and the highly optimized DeepSeek-V4-Flash (284B total / 13B activated parameters).Pre-trained on over 32 trillion tokens, the series incorporates breakthrough hybrid attention mechanics (CSA/HCA) and the Muon optimizer. This drastically reduces hardware demands, requiring just 10% of the KV cache compared to previous generations, making complex, long-horizon tasks and test-time scaling routine.Links: Access the model checkpoints on the official Hugging Face Collection.Would you like a bulleted key-features list to go right next to this text, or do you need a version under 50 words?Try without personalization