DeepSeek has released DeepSeek-V4, a new open model family designed to address the computational bottlenecks of running frontier models as agents over million-token contexts. The architecture introduces Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to drastically reduce single-token inference FLOPs and KV cache memory requirements.
- V4-Pro requires 27% of the single-token inference FLOPs and uses 10% of the KV cache memory compared to DeepSeek-V3.2, while V4-Flash reduces these figures to 10% and 7% respectively.
- The model preserves reasoning traces across user message boundaries when tool calls are involved, enabling coherent long-horizon agentic workflows.
- A new XML-based tool-call schema with dedicated tokens aims to reduce parsing failures associated with nested JSON content.
- Agent behavior was trained using DeepSeek Elastic Compute (DSec), a Rust-based sandbox infrastructure supporting concurrent RL rollouts across various execution substrates.
- V4-Pro-Max achieved 80.6 on SWE Verified and 67.9 on Terminal Bench 2.0, placing it competitive with leading proprietary models like Opus 4.6-Max and GPT-5.4-xHigh.
The release provides four checkpoints (two instruct, two base) for DeepSeek-V4-Pro and V4-Flash, offering the community a cost-efficient alternative for long-context agent applications.