Hybrid LLM (Gated DeltaNet)

A hybrid model mixes Gated DeltaNet layers (linear attention with a gated, delta-rule state: memory and time grow linearly with the input) with a few full-attention layers. The catalog entry is hybrid (Qwen3.5-0.8B, Qwen3.5-2B; the vision tower stays frozen); every LLM and decision recipe accepts it, and LoRA adapts the linear-attention projections too. On a Mac without the fused kernels, its linear-attention layers run a slower fallback.

Experiment	Role of the hybrid model	Result
Jarvis inner-thought router	the backbone compared with Qwen3-0.6B	166 of 167 held-out, ECE 0.006, 0.30 s a request; not chosen, because its one boundary mistake was sure (0.997) and it was four times slower

No experiment ships a hybrid model yet; the first one gets its page here.

Source: handwritten. This page is generated; edit the source and run npm run sync.