[techmore]

What

The Apple-native half of a same-machine comparison: every model × length × load from the M1 Pro scoreboard, run under macOS 27.0 (26A5421a) on the Mac mini M4 (Mac16,10, 10-core GPU, 16 GB) with the repo's own engine-compare.py, so the same commands can later run under Linux on this mini and compare row for row.

• engine-compare.py: macOS arms mlx-metal (in-process, same worker), lm-metal (stock mlx_lm.server) and tensorfold-metal (TensorFold v1.1.9 release, pinned sha256, --backend mlx --no-drafts). Models come from the HF cache at the pinned revisions (weights hash-checked as before). New env knobs also let the Linux side on the mini use HF snapshots and the packaged venv: MSERIES_HF_SNAPSHOTS, MSERIES_MLX_PY, MSERIES_TF_VULKAN_BIN, MSERIES_LFM3_MODULE, plus --reserve-gib.
• machine.py: macOS memory (kern.memorystatus_level) and thermal (pmset -g therm) guard; receipts record load, top CPU users, limactl list, power and swap at start and end.
• scoreboard.py: a separate M4 section, engine-scoreboard-m4-mini-macos.csv / .svg, never ranked against the M1 Pro, with a "Same output as M1 Pro (Linux)" column and a "How to run the same on Linux on the M4 mini" note (points at omarchy-mlx m4-mini-bench / omarchy-mlx-m4-bench). The M1 Pro table regenerates byte-for-byte unchanged.
• Receipts: receipts/2026-10-31-m4-mini-macos/ (8 runs, the three macOS legs of m4_mini_bench.py, venv freeze, README).

Pins: Python 3.14.3, mlx 0.32.3, mlx-lm 0.31.3, transformers 5.16.1 (the pins m4-mini-bench names for macOS).

Results (median of 3, 32 tokens unless noted)

Model	In-process MLX (Metal)	mlx_lm.server	TensorFold Metal
Granite 4.0 1B 4-bit	58.6 tok/s, TTFT 0.22 s	43.0, 0.22 s	–
LFM3-1.7B 8-bit	31.5, 0.24 s	–	–
Qwen3.5-2B 4-bit	47.7, 0.38 s	42.4, 0.42 s	39.7, 0.07 s
Qwen3.5-2B, 256 tokens	51.3, 0.52 s	40.6, 0.55 s	36.4, 0.12 s
Qwen3.5-2B, 4 / 8 at once (total)	–	41.6 / failed	36.8 / 35.4
Qwen3.5-4B 4-bit (4 at once)	29.0, 0.39 s	25.2 (29.2)	–

Outputs match the M1 Pro's Linux text except Qwen3.5-2B at 256 tokens (MLX/server on Metal diverge after 1,207 of 1,301 chars; TensorFold matches) and Qwen3.5-4B in-process (after 55 chars; the server matches). mlx_lm.server at 8 simultaneous requests had 1–2 connections reset each round (listen backlog of 5 on macOS); reported under the table.

Caveat: the Lima kbuild CI VM was busy throughout (390–590 % CPU, load average 15–23, 10 of 16 GB held). It was not touched; these are a loaded-machine floor. Compare against a Linux run under similar load, or re-run both quiet.

🤖 Generated with Claude Code

[techmore]

Superseded by the archived ElementaryOS experiment and the active Aurora M4 repositories.