DeepSeek-V4-Flash and V4-Pro released as open-weight models with hybrid attention and 13B active params

DeepSeek released V4-Flash (284B total / 13B active parameters) and V4-Pro as open-weight models on Hugging Face. Both are optimized for agentic coding and long-context efficiency, with structural innovations including hybrid attention and DeepSeek Sparse Attention to reduce compute and memory costs. V4-Flash is positioned as a cost-effective option for developers building long-context and agentic workflows.