DeepSeek-V4: a million-token context that agents can actually use

DeepSeek’s new V4 release introduces a 1‑million‑token context window across two checkpoints: DeepSeek‑V4‑Pro, a 1.6‑trillion‑parameter model with 49 billion active parameters, and DeepSeek‑V4‑Flash, a 284‑billion‑parameter model with 13 billion active parameters. The architecture alternates between Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) layers, achieving a 27 % reduction in single‑token inference FLOPs compared to DeepSeek‑V3.2 and cutting KV cache usage to 10 % for the Pro model and 7 % for Flash. These efficiencies stem from block‑level compression (4× in CSA, 128× in HCA) and the use of FP8 storage for most KV entries. Post‑training, the model retains reasoning traces across tool‑call rounds and user turns, enabling coherent, cumulative chains of thought for multi‑turn agent tasks. It also introduces a |DSML| special token and an XML‑based tool‑call format to reduce parsing failures common with JSON. The model was trained with reinforcement learning in a Rust‑based sandbox platform, DeepSeek Elastic Compute, which supports function calls, containers, microVMs, and full VMs via a unified Python SDK. In internal benchmarks, V4‑Pro‑Max achieved a 67 % pass rate on 30 curated coding tasks, outperforming Sonnet 4.5 and matching Opus 4.5. Retrieval tests show MRCR 8‑needle accuracy above 0.82 up to 256 K tokens and 0.59 at 1 M tokens. The release positions DeepSeek V4 as a competitive open‑source alternative for agentic workloads, with its efficient long‑context handling and tooling support potentially influencing how the community builds and deploys AI agents.