Machine
Apple M2 Ultra · Apple M2 Ultra #c9279b01
22 runs of 6 models since Aug 31, 2026. Last run Sep 4, 2026.
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
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
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.
configrunsqualityeach runspeed
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
gemma-4-31B-it-MLX-8bit
1 run · 1 config
gemma-4-26B-A4B-it-MLX-8bit
1 run · 1 config
gemma-4-12B-it-bf16
1 run · 1 config
Compare with another machine
Run history
Sep 3, 2026
23:53Qwen3.8-27B-oQ8e-fp16-mtpoMLX31.5 tok/s81.623:25Qwen3.8-27B-oQ8e-fp16-mtp×5oMLX35.3 tok/s83.722:36Qwen3.8-27B-oQ8e-fp16-mtp×2oMLX34.3 tok/s67.2