EN
11 November 2026 · Jakub M.
Two identical Mac Studio M5 Ultra machines: does local AI speed differ?
We have two Mac Studio computers with the M5 Ultra and 256 GB of unified memory: one has been running since October, the other is brand new. We ran the same test with three language models on both. Generation speed differed by less than half a percent.
Two computers of the same model and configuration should be equally fast. Small differences between individual units are sometimes reported, though, and we wanted to know how big they are for local AI. For us it is a practical question: a customer rents a configuration and should get the same performance whichever unit they are given.
What we measured on, and how
- Hardware: two Mac Studio computers with the Apple M5 Ultra (36-core CPU, 80-core GPU) and 256 GB of unified memory. One has been in use since October, the other is new.
- Software: identical on both: macOS 27.2.1, Python 3.12, MLX 0.32.4 and mlx-lm 0.32.2.
- Models: 4-bit MLX versions from the mlx-community organisation on Hugging Face: Qwen3 30B-A3B, Llama 3.3 70B Instruct and gpt-oss-120b (MXFP4-Q8).
- Test: the built-in
mlx_lm.benchmarktool with a short prompt (512 tokens, three runs) and a long prompt (8,192 tokens, two runs), each followed by 128 generated tokens. The prompt consists of random tokens with a fixed seed, so both Macs received exactly the same input. Every measurement started with a warm-up run, and both Macs were measured on the same day.
Results
| Model | Prompt (tokens) | Prompt processing, older Mac | Prompt processing, newer Mac | Generation, older Mac | Generation, newer Mac | Peak memory |
|---|---|---|---|---|---|---|
| Qwen3 30B-A3B | 512 | 2,285 | 2,237 | 162.2 | 162.2 | 17.8 GB |
| Qwen3 30B-A3B | 8,192 | 7,646 | 7,602 | 138.1 | 138.0 | 18.7 GB |
| Llama 3.3 70B | 512 | 469 | 467 | 25.2 | 25.3 | 40.3 GB |
| Llama 3.3 70B | 8,192 | 682 | 683 | 23.6 | 23.6 | 43.3 GB |
| gpt-oss-120b | 512 | 1,077 | 1,046 | 121.4 | 121.2 | 63.8 GB |
| gpt-oss-120b | 8,192 | 5,143 | 5,109 | 114.8 | 115.2 | 64.5 GB |
Prompt processing and generation in tokens per second, average of the runs (three for 512 tokens, two for 8,192), 128 generated tokens. Older Mac: in use since October; newer Mac: brand new. Same configuration, 4-bit MLX models, mlx-lm 0.32.2, macOS 27.2.1, measured on 11 November 2026.
What the numbers show
Generation speed, that is how fast the model writes its answer, is the same on both Macs: across all three models and both prompt lengths it differs by at most 0.35 %. Processing the long prompt differs by at most 0.7 %. Peak memory use is identical to a tenth of a gigabyte.
With the short prompt, the new Mac came out 2 to 3 % slower at processing it. That run takes only 0.2 to 0.5 seconds, however, and repeated runs on one and the same Mac varied by up to 3 %. We therefore do not consider this difference real.
The results also match our October measurement: gpt-oss-120b generated 120 tokens per second then and 121 now.
Limits of the test
We compared only two units, so the result does not speak for every Mac Studio M5 Ultra. A prompt of random tokens and short generation give pure speed figures; with real tasks and longer answers the speed may differ. The test does not measure the quality of the answers.
Try it yourself
Our RAM calculator shows how much memory the model you want needs. To run your own tests, you can rent a Mac Studio M5 Ultra with 256 GB from one day.