M5 Ultra Mac Studio vs my four DGX Sparks: the same model
MacStories tested GLM 5.3 Flash, the model my cluster runs, on a 256GB M5 Ultra Mac Studio, in a review I had missed. One Mac Studio writes about as fast as my four DGX Sparks and my Sparks read a prompt 2.04 times as fast, but the comparison is a rough one.
Plays here, from YouTube
This story starts with a correction. In episode 6 I said nobody had measured one of Apple's new Mac Studios. That was wrong, because MacStories had published a detailed review on 21 September, a week before my video, and I missed it. So I have put their numbers next to mine.
| Measure | M5 Ultra Mac Studio, 256GB | My four DGX Sparks, 512GB |
|---|---|---|
| Writing | 41 tok/s | 33.8 tok/s on prose, 58.4 tok/s on code |
| Reading a prompt | 1,107 tok/s on 13,639 tokens | 2,257 tok/s on about 15,500 tokens |
| The model's thinking | Counted in the writing speed | Switched off |
The Mac Studio's figures are from MacStories' review of 21 September 2026. The DGX Sparks' figures are my own measurements of 2 October 2026.
Writing
They tested the 256GB M5 Ultra, and one of the models they ran was GLM 5.3 Flash, the same model that is on my cluster. In their test the Studio wrote at 41 tokens a second. When I measured my four Sparks on Friday 2 October, they came out at 33.8 on prose and 58.4 on code, taking the middle of six runs. So one Mac Studio writes about as fast as my whole cluster.
Reading the prompt
Reading the prompt is where the two differ. Their Studio read a 13,639-token prompt, which the review labels 16K, at 1,107 tokens a second. My four Sparks read one of about 15,500 tokens at 2,257 tokens a second, the middle of three passes. By my own maths that is 2.04 times as fast (2,257 divided by 1,107). That fits what I said in episode 6, that reading a prompt spreads across boxes much better than writing does.
A rough comparison
I should say the comparison is a rough one.
- The builds and the software are different. MacStories ran a mixed 4 and 8-bit MLX build of the model.
- Their writing speed includes the model's thinking, where mine had thinking switched off.
- With a very long prompt of 258,025 tokens their Studio ran out of memory. My cluster has twice as much (512GB against 256GB). Apple's store page said the M5 Ultra's 512GB memory option would arrive in late October.
- On 3 October I still could not find an independent test of several Studios linked together.
Read more
- MacStories' review of the M5 Ultra Mac Studiomacstories.net
- Apple's Mac Studio store page, with the 512GB option listed as comingapple.com
The numbers in this story
| Figure | What it is | From | Kind |
|---|---|---|---|
| 41 tok/s | GLM 5.3 Flash writing on the 256GB M5 Ultra Mac Studio | macstories.net | a reviewer's test |
| 33.8 tok/s on prose, 58.4 on code | The same model writing on my four DGX Sparks, the middle of six runs | my own test | measured by me |
| 1,107 tok/s on a 13,639-token prompt | The Mac Studio reading a prompt | macstories.net | a reviewer's test |
| 2,257 tok/s, 2.04 times | My four DGX Sparks reading a prompt, against the Mac Studio | 2,257 divided by 1,107 | my own maths |
| No answer | The Mac Studio on a 258,025-token prompt | macstories.net | a reviewer's test |
My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.
If a number on this page is wrong, I would rather know. Tell me in the comments under the video and I will correct it here and pin the correction there. What has been corrected so far is on the corrections page.