Machine

Apple M2 Ultra · Apple M2 Ultra #c9279b01

22 runs of 6 models since Aug 31, 2026. Last run Sep 4, 2026.

ShareRedditX
22
runs
6
models
10
configs
5
configs run 2+ times

Hardware

GPU
Apple M2 Ultra 128 GB
CPU
Apple M2 Ultra 24 cores
RAM
128 GB
OS
macOS 26.6.2
Tools
oMLX 22
Client
0.4.53
All Apple M2 Ultra benchmarks

Models on this machine

One dot per model, at its best config. Up is better answers, right is faster. Time to finish counts every token the model wrote, thinking included, so a model that talks a lot ends up further left.

Best trade-off: quality 60+, and nothing else here is both faster and betterOther models · dot size = number of runs
below 20 tok/s: slower than you read40506070809010020tok/s (log scale) →quality →outliers sit at the edges

Medians of each model's best config, 6 models.

Best config per model

sort

Runs with the same model, quant, tool, context, thinking and KV cache count as one config. Numbers are the median of its runs, and each dot is one run on the 0–100 quality scale. ★ marks the best config with at least 2 rated runs.

Qwen3.8-27B-oQ8e-fp16-mtp
13 runs · 4 configs
★ bestFP16oMLX256k ctxthinking on
×4 runs
83.981.6–85.6
steady
34.2 tok/s
29.8–34.9
Qwen3.8-Flash-Next-oQ4e-mtp
5 runs · 2 configs
★ bestoMLX128k ctxthinking on
×4 runs
81.478.0–84.4
steady
25.9 tok/s
23.8–27.2
Qwen3.6-27B-oQ8-mtp
1 run · 1 config
★ bestoMLX256k ctxthinking on
79.8
1 rated run
32.8 tok/s
gemma-4-31B-it-MLX-8bit
1 run · 1 config
★ bestoMLX256k ctxthinking on
74.1
1 rated run
16.6 tok/s
gemma-4-26B-A4B-it-MLX-8bit
1 run · 1 config
★ bestoMLX256k ctxthinking on
73.4
1 rated run
68.8 tok/s
gemma-4-12B-it-bf16
1 run · 1 config
★ bestoMLX128k ctxthinking on
66.4
1 rated run
24.0 tok/s

Compare with another machine

Run history