Atlas Reference Setup open-sources Mac mini M4 assistant

The honest lesson here is that local-first only works when you stop treating the Mac mini like a universal substitute for frontier cloud models. The repo is more convincing as an ops/security playbook than as a model benchmark.

• –Local Whisper is the clearest keep-local win: strong quality, low latency for async voice, and a real privacy benefit.
• –Qwen 3.5 27B as offline fallback makes sense on 24GB, but it’s a continuity layer, not the main reasoning engine.
• –Piper TTS is a practical compromise, while FLUX.1-schnell on MPS sounds like the first thing I’d keep out of the critical path.
• –The security scaffolding matters as much as model choice: allowlists, sandboxing, integrity checks, and cron watchdogs are what make 24/7 assistants believable.
• –The pattern worth copying is the split itself: cloud for heavy reasoning, local for sensitive, frequent, and failure-prone tasks.