A small local mind for the valley
Four small Qwen and Gemma candidates could power residents inside the offline game. The shortlist compares vision support and actual bundle sizes; the next step is to test their choices while the valley is running.
Full entry & measurements
Research only: docs/local-model-shortlist.md records current Qwen3.5-2B/4B and Gemma 4 E2B/E4B, their Apache 2.0 weight licenses, and documented local vision support. Qwen3.5-4B is the main quality candidate, Gemma 4 E2B the Gemma challenger, and Qwen3.5-2B the smaller-footprint candidate; these are proposed evaluation roles, not measured winners. Gemma effective parameter labels differ from total weight size. Measured study scope on 2026-10-15T00:08:57.138404+02:00, source model-study-a749f8ca41a44279: 4 primary candidates reviewed; 4 selected model-plus-projector bundle sizes checked from public Hugging Face metadata at pinned revisions; 0 new inference benchmarks. Bundles range from 1,949,063,104 to 6,146,493,536 bytes; these are file sizes, not peak RAM. Exact byte counts, revisions and links are in the source document. Existing Qwen3-VL-2B live evidence remains historical. Local-first delivery, a fully offline installer and continuous local play are selected planning directions; packaging and quota changes are not implemented here. No weights were downloaded and no resident profile or runtime state changed. The reused 14 October garden image illustrates the visual environment, not a new model test. Next: compare valid intents, coherent conversations, privacy, memory, latency and Unity frame times with identical fixtures, followed by a 30-minute offline acceptance run.