Decoder LLM
A decoder-only transformer reads its input with full attention in every layer. The catalog entry is transformer
(Qwen3-0.6B, Qwen3-1.7B); every LLM and decision recipe accepts it. At small sizes it is the fastest option on a Mac:
in the router experiment Qwen3-0.6B answered in 0.08 s against 0.30 s for the hybrid Qwen3.5-0.8B.
| Experiment | Recipe | Model | Result |
|---|---|---|---|
| Jarvis inner-thought router | llm_decision_sft, pointer head | Qwen3-0.6B | 166 of 167 held-out, ECE 0.016, 0.08 s a request |
Source: handwritten. This page is generated; edit the source and run npm run sync.