This page details the optional education card on Roxy’s status page. It is a teaching lab on Bob’s desk, not a second production brain and not tomorrow’s session work.
Status: optional plan only. Do not install it as part of proving cited answers or building company-event ingestion. Roxy on Nous Hermes Cloud stays the production assistant. LM Studio is the easy tool — not the fastest inference engine.
Bob has a 64 GB Mac Mini. A 27-billion-parameter local model fits it well and is a good way to see how on-device models behave — speed, context limits, when thinking helps, when it just burns tokens.
It is not a replacement for Cloud Roxy. The library, mail path, Honcho, and cited answers stay on Hermes Cloud. The Mini is a lamp you switch on, not a server that runs all day.
| Item | Choice | Why |
|---|---|---|
| App | LM Studio, Apple Silicon build | Easy UI, local OpenAI-compatible server, MLX runtime |
| Weights now | lmstudio-community/Qwen3.6-27B-MLX-8bit (~29.5 GB) | Known-good Mac quant. Leaves headroom for macOS, chat, and context |
| Weights later | Official Qwen/Qwen3.8-27B, then LM Studio’s MLX 8-bit | 3.8 is not released. Community 3.8 repos today are placeholders — do not download them |
| Not first | BF16, 4-bit, or random GGUF forks | BF16 is too tight on 64 GB; 4-bit wastes the RAM we have; GGUF is slower on this Mac |
LM Studio is the on-ramp. A later, faster path on the same machine would be mlx_lm.server or rapid-mlx. That is not required for the education goal.
127.0.0.1 only unless a later, explicit LAN decision says otherwise.http://127.0.0.1:<lms-port>/v1.