DeepSeek AI Releases DeepSeek-V4: Compressed Sparse Attention and Highly Compressed Attention Enable Million Token Content
DeepSeek-AI has released a preview version of the DeepSeek-V4 series: two Mixture-of-Experts (MoE) languages designed for the single main challenge of making million-token windows usable and affordable during speculation. This series consists of DeepSeek-V4-Pro, with 1.6T total parameters and 49B activated per token, and DeepSeek-V4-Flash, with 284B total parameters and 13B activated per token. Both … Read more