Local models on a Mac mini M4 32 GB — benchmark log

Field notes from six months of running local LLMs, image generators, and a music-analysis model on a Mac mini M4 with 32 GB unified memory.

📖 Read the report: https://rachokthebot-dev.github.io/local-models-blog/ 🔬 Bench scripts + raw data: bench/

What's in here

• index.html — the full benchmark log, self-contained, no build step. Served via GitHub Pages.
• bench/ — every script that produced a number on the page, plus the raw *.json outputs and markdown reports.

Sections of the report

1. Executive summary — what we kept
2. The hardware and the constraints
3. General-purpose LLMs (reasoning · vision · long-doc · tool-call)
4. Image generation (FLUX.2 Klein 4B / 9B)
5. Music analysis (SongFormer beats audio LLMs)
6. Models on the watchlist
7. The performance ceiling, explained
8. Methodology and caveats

Reproducing the benchmarks

See bench/README.md for the run-it-yourself commands. Hardware target: Mac mini M4, 32 GB; runtimes: rapid-mlx 0.4.2 + Ollama 0.30.4 (MLX backend).