# Hybrid Linear-Attention Architectures

Open models will keep replacing much of full attention with fixed-size recurrent layers such as Gated DeltaNet, and that fixed state will increasingly be used as an engineered resource for memory, tool execution and cheaper long-context inference.

- Conviction: 40 / 100 (forming)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-10-29T00:00:00.000Z
- Last updated: 2026-10-29T06:40:54.498Z
- Canonical: https://polylog.news/ai/trends/hybrid-linear-attention-architectures
- Publisher: Polylog

## Recent evidence

- [confirms] Developer compiles an exact RISC-V CPU into Qwen3.5-4B and uses its Gated DeltaNet state as memory (2026-10-29): A developer's Qwenputer compiles an exact RISC-V CPU into Qwen3.5-4B and uses its Gated DeltaNet recurrent state as working memory, running C and MicroPython inside the forward pass and scoring 97.5% to 100% on DeepMind arithmetic tasks. This shows the fixed recurrent state being used as an engineered resource for tool execution, at the cost of doubling the checkpoint to 18.8 GB.