# Mote local-model trial notes — October 6, 2026

Author: Nick Bryant. Publication summary of the original project notebook; not raw inference logs or an independently reproduced benchmark.

Hardware: Apple M2 Max, 32 GB unified memory. Synthetic isolated companion sessions. Model labels are reproduced as recorded in the notebook; exact model digests and a complete sampling configuration were not preserved in this summary, so these notes are insufficient for exact reproduction.

| Recorded model label | Output mode | Samples | Valid structured decisions | Recorded response latency | Observation |
|---|---|---:|---:|---|---|
| Gemma 4B | Unconstrained JSON | 6 | 2/6 | Approximately 2–3.5 seconds on valid replies | Required memoryIds frequently omitted |
| Gemma 4B | Required JSON schema | 6 | 6/6 | Median 2.64 seconds; max 3.47 seconds | Repetition and scene embellishment remained |
| Qwen 0.6B | Required JSON schema | 1 | 1/1 | 1.10 seconds | Echoed input |
| Gemma 1B | Required JSON schema | 1 | 1/1 | 2.11 seconds | Produced a companion reply |

A further isolated packaged-app greeting took 2.76 seconds including 1.60 seconds loading, and echoed part of the request. The notebook also records later relevance-gated turns of 1.03 and 0.77 seconds, with conversational limitations still present. These are different short checks, not pooled comparative measurements.

Initial samples used a 160-token output limit; the implementation later raised it to 256 because the smaller limit could truncate the JSON envelope. Background model calls were disabled by default. No all-day thermal or battery measurement was performed. Node process memory excludes the model process. No claim of consciousness or human equivalence follows from these observations.

The next useful study should fix exact model versions, scenario scripts, sampling settings, warm/cold load conditions, and cadence; preserve outputs; measure whole-machine resource use separately; and rate useful conversation independently of schema validity.