the find
− 0.6 mThe smallest project in the catalog: two shell scripts that launch local LLM brains over MLX for the avatar stack. But the comments are the artifact — a densely annotated field essay on running large models for live speech, where every token of latency is a beat of dead air on a call.
cross-section
− 1.3 mThe scripts serve two candidate brains from the laptop, bound to the private mesh only — never a public interface. The essay's first law: every reasoning model in the stack must have thinking disabled. A reasoner given a greeting will happily spend its token budget deliberating; in text that's a delay, in speech it's silence on the line — and sometimes, after all that deliberation, it returned no content at all.
inscription
− 2.0 mField note, dated 2026-07-28, from the comments: Gemma once spent 108 tokens reasoning about how to say "hi." The same day's entry records why the 26B model beat a 120B model as the resident brain at identical latency: 26 GB versus 59 GB of weights on unified memory, and the smaller footprint stops the brain from starving the face renderer sharing the same RAM. It also simply played the part better — it remembered a temple visit the bigger model denied ever happened, and kept the correct Thai speech register throughout. Capability benchmarks measure the wrong thing; the right question was who could hold a conversation while leaving room for a face.