Heavily Compressed Attention

A DeepSeek-V4 long-context attention path that relies on a more aggressively compressed memory view when full-detail history is too costly.

Compressed sparse attention still keeps a route back to selected high-resolution detail. Heavily compressed attention goes further toward a compact memory view so very long contexts remain feasible even when detailed revisit must be rare.

At a glance

Released

June 2026

Authors

DeepSeek-AI

Optimizes

  • Kv Cache
  • Long Context Inference
  • Memory Bandwidth

What It Is

Heavily compressed attention is the more aggressively compressed side of the DeepSeek-V4 hybrid attention story. Instead of preserving broad access to detailed tokens, it leans harder on a compact representation of old context.

Why It Exists

The module targets extreme long-context efficiency, especially KV-cache growth and memory bandwidth pressure.

How It Works

Older context is pushed into a more compressed memory structure, and queries interact with that compact summary rather than reopening a large detailed history by default.

Math Or Compute Schema

The sketch below keeps the central mental model simple: the history is compressed harder, then queried through that compact path.

Compared To Nearby Modules

CSA preserves a compressed path plus selective detailed retrieval. HCA moves farther toward compression-first memory, which can be cheaper but also more lossy.

Example Architectures

DeepSeek-V4 presents HCA alongside CSA as part of its hybrid attention design for million-token context.

Limitations And Tradeoffs

Heavier compression risks hiding details that would matter if the summary path is too coarse. The gain is efficiency, but the tradeoff is a less faithful memory of distant exact tokens.

Why It Still Matters

It is a useful reference because it shows that long-context design is not one knob. Systems can mix multiple memory paths with different cost and fidelity.

Tags

References

  1. DeepSeek-AI. "DeepSeek-V4 Technical Report." 2026.