Two 4B-class models trade activation vectors 8x per token by injecting into each other's residual streams
Sauers_ · x · 2026-10-29
Arcee showed an unusual experiment: Quill (Qwen3-4B) and Ember (AFM-4.5B) exchange activation vectors 8 times per token, with layer outputs injected directly into each other's residual stream so each model can attend to its partner's past tokens — a novel form of cross-model collaborative inference.
More from Models
- Persona prompt pins models at ~35% on SIMMER benchmark regardless of Claude or DeepSeek — amplifiedamp · 2026-10-29
- OpenAI Terra 6.1 in practice: cheap model plus tiered follow-up beats what benchmarks miss — IamNotGorbachev · 2026-10-29
- qwen-code PR Sizes All Qwen4 Models at 1M-Token Context Window — haider1 · 2026-10-29
- Developer flags Claude weekly limit showing 2% used with zero activity after reset — umeshonai · 2026-10-29
- 'Persona neurons' discourse overhyped, but tiny perturbations breaking alignment is telling — aiamblichus · 2026-10-29
- LocalLLaMA users draft a Qwen4 wishlist: 512K context and plainer answers — perelmanych · 2026-10-29