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

DeepSeek-V4 introduces a million-token context window optimized for agent workloads, tackling the KV cache problem with a hybrid attention design (CSA and HCA) that alternates compressed and dense attention to reduce FLOPs and memory. It supports long-running tool-use trajectories and includes an agent-specific post-training recipe, yielding faster inference and far lower cache usage versus prior architectures.

Source

HuggingFace Blog

Ingéré

24 AVR. 2026 · 19:10

Score édito

5.0 / 5