Ollama's MLX builds of Gemma 4 are ~1.2x faster on a M2 Pro MacBook

This is my personal blog, where I write solely in my personal capacity. It does not represent the positions of any organisations I'm associated with.

Also, everything here is provded "as is", without warranties of any kind, express or implied.

9 August 20269 August 2026

Ollama now ships MLX builds of Gemma 4 with multi-token prediction.

On my M2 Pro MacBook with 16GB RAM, the -mlx tags are ~1.2x faster than the standard ones, for the same model:

The MLX builds are also smaller on disk (e2b: 6.5GB vs 7.2GB, e4b: 8.8GB vs 9.6GB), and produced the same answers in my spot checks. So on Apple Silicon there doesn't seem to be much reason not to use them.

Raw data

Median of three runs generating 1000 tokens each, after 60 seconds of warm-up:

Model	Standard tps	MLX tps	Speedup
gemma4:e2b	61.8	74.2	1.20x
gemma4:e4b	32.6	37.9	1.16x

Individual runs:

Model	Run 1	Run 2	Run 3
gemma4:e2b	61.3	61.8	62.6
gemma4:e2b-mlx	74.2	73.6	74.8
gemma4:e4b	30.9	32.6	32.6
gemma4:e4b-mlx	36.7	37.9	38.9

Run with ollama 0.32.4 on 2026-08-09, at default settings for each model tag.

Appendix: comparison to April 2026 results

These absolute numbers are a bit lower than the 66.3 tps for gemma4:e2b and 41.3 tps for gemma4:e4b I measured in April, on the same machine.

I wondered whether ollama had regressed (the April run was with 0.20.7 and this is 0.32.4). After re-downloading 0.20.7, I found gemma4:e2b ran at 59.5 tps, versus 61.8 tps on 0.32.4. So there appears to be no regression here; if anything the newer version is very slightly faster.

The remaining gap is probably just my laptop being warmer in the London heatwave than it was in spring.

Either way, each standard/MLX pair was measured back to back in one session, so the relative difference probably holds even if absolute numbers drift between sessions.