Mac mini tested as an Agent server: smooth local 27B models, 70B unusable
APPSO · wechat · 2026-09-21
APPSO ran an in-depth hands-on with the new Mac mini as an "AI PC," concluding it works best as an always-on agent server rather than a heavy local-model workstation.
- Local model benchmarks: a 4-bit 27B Qwen model summarizes long docs at 32 token/s; 70B Llama 3.1 is effectively unusable (a single reply took 3 minutes). On a 48GB M5 Pro, the sweet spot is 20-30B models. GGUF is cross-platform; MLX is built for Apple Silicon's unified memory — which is faster depends on the model and runtime.
- Agent workflows: LM Studio Bionic detects device memory and flags "Full GPU Offload Possible" vs "Likely too large," and can read files, run commands, and execute multi-step tasks. In tests, an agent drove Blender via its Python API to build and render a 3D model, then made a landing page. Video generation is heavy: a 15s clip took nearly 90 minutes.
- New usage pattern: the author used the Mac mini as a Codex agent server, with four agents simultaneously writing an Xcode app, compiling Chromium, processing 100GB of media, and rendering in Blender — while the human kept using the machine. The classic "multitasking" PC is turning into a "multi-agent" machine, and a small always-on box is arguably the earliest PC form factor to fit that shift.
More from coding & agent
- User automates entire video clipping workflow with Nous Research's Hermes agent — Teknium · 2026-11-12
- As models improve, writing good evals gets harder and more important — amankhan · 2026-11-12
- Prime agent ships with day-one multi-agent communication, runs in Minecraft unmodified — sarahcat21 · 2026-11-12
- Autonomous agent edits just 1,024 weights to lift Qwen3.5 0.8B from 7.4% to 37.4% on MATH-500 — VictorSanh · 2026-11-12
- AI game jam with Anthropic, Google, OpenAI teams: cash prizes plus $50K continuation grants — chrisfirst · 2026-11-12
- Free AI game jam in SF Nov 14–15 with teams from Anthropic, Google, Meshy, OpenAI — chrisfirst · 2026-11-12