Attention

Linear attention (gated DeltaNet)

Replaces the attention matrix with a fixed-size recurrent state - cost stays constant per token, no KV cache growth.

Linear attention keeps a compact running state that is updated once per token (here with the gated delta rule: the state is selectively decayed and rewritten), so generation cost does not grow with context length and there is no per-token KV cache. Qwen3.5-style hybrids use linear attention in most layers and keep periodic full-attention layers for precise long-range recall - the best of both worlds for long contexts.

See it in real models

Open any of these on hfviewer to find this block in the interactive architecture graph.

Related concepts

hfviewer renders the full architecture of 2,300+ Hugging Face models as interactive graphs - hover any block to see what it does, with this model's real numbers.

Browse all model graphs →