DeepSeek-V4: a million-token context that agents can use (huggingface.co)

DeepSeek has announced the release of DeepSeek-V4, a major upgrade designed to handle long-context agent workflows more efficiently with a revolutionary one-million-token context window. This new architecture addresses known issues from previous models, such as context budget limitations and degradation of performance during extended tool-use tasks. Key improvements include a 27% reduction in single-token inference FLOPs compared to its predecessor, DeepSeek-V3.2, and an impressively low 2% of the KV cache size needed, making it significantly more deployable for applications involving large context handling. The enhancements come from novel mechanisms like Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), which alternate across layers to optimize attention patterns, supporting effective long-context inference. The model also maintains reasoning across multiple user interactions, ensuring continuous thought processes during multi-turn dialogues. Furthermore, the introduction of a new XML-based tool-call format minimizes parsing errors often seen with JSON, enhancing reliability in agentic tasks. With competitive benchmarks showcasing its capabilities, DeepSeek-V4 sets a new standard for open models in AI, potentially reshaping deployment in real-world scenarios, while challenging the community to build compatible tool harnesses.