EN
27 October 2026 · Jakub M.
Clash of the titans: MacBook Pro M5 Max vs Mac Studio M5 Ultra in local AI
The fastest MacBook against the fastest Mac Studio, 32 language models, one winner. In the Vše Smart test the Mac Studio M5 Ultra was faster in 31 of 32 rounds, by a median of 89 percent. With 70B models and larger it was more than two and a half times faster.
Source: MacBook Pro M5 Max vs. Mac Studio M5 Ultra: srovnání výkon lokálních AI modelů, Ludvík Eckerman, Vše Smart, 7 October 2026 (in Czech). All figures below come from that article.
Apple’s most powerful laptop against its most powerful desktop. Both machines in their top configurations, the same models, the same prompt, the same room. The Czech site Vše Smart put them through 32 local language models, and the result is a clear win on points with a few surprising rounds.
The fighters
| MacBook Pro M5 Max | Mac Studio M5 Ultra | |
|---|---|---|
| CPU cores | 18 | 36 |
| GPU cores | 40 | 80 |
| Unified memory | 128 GB | 256 GB |
| SSD | 4 TB | 4 TB |
On paper the M5 Ultra is two M5 Max chips joined together: twice the CPU and GPU cores and twice the memory. The question is how much of that turns into real speed.
The rules
- Models ran locally in Ollama, no cloud.
- One prompt for all: a long technical text on modern car design, up to 2,000 tokens, 8k context, temperature 0, fixed seed, thinking mode off.
- Each model ran 11 times; the first run was discarded and the result is the average of the remaining ten.
- Models in random order, with a 30-second pause after each, so heat would not favour either side.
All 32 rounds
| Model | MacBook Pro M5 Max | Mac Studio M5 Ultra | Studio vs MacBook |
|---|---|---|---|
| qwen3:4b | 121.39 | 203.31 | +67.5% |
| qwen3:30b | 112.63 | 170.12 | +51.0% |
| gpt-oss:20b | 100.94 | 169.35 | +67.8% |
| qwen2.5:7b | 81.36 | 154.19 | +89.5% |
| qwen3.5:35b-mlx | 124.13 | 144.87 | +16.7% |
| qwen3.5:9b | 68.06 | 126.59 | +86.0% |
| gpt-oss:120b | 73.15 | 115.21 | +57.5% |
| gemma4:26b | 68.07 | 111.74 | +64.2% |
| qwen3-coder-next:latest | 75.91 | 109.48 | +44.2% |
| qwen3-next:80b | 73.14 | 94.48 | +29.2% |
| gemma4:26b-mlx-bf16 | 58.76 | 93.21 | +58.6% |
| gemma4:e4b | 93.12 | 92.95 | −0.2% |
| qwen3.8:27b-mlx | 37.90 | 90.65 | +139.2% |
| gemma4:12b | 46.62 | 90.48 | +94.1% |
| ministral-3:14b | 44.61 | 88.66 | +98.7% |
| gemma3:12b | 47.00 | 88.26 | +87.8% |
| qwen3.5:122b-a10b | 46.91 | 84.80 | +80.8% |
| phi4-reasoning:plus | 36.83 | 76.10 | +106.6% |
| llama4:scout | 34.12 | 63.26 | +85.4% |
| mistral-small3.2:24b | 26.42 | 62.19 | +135.4% |
| qwen3.8-flash-next:125b-mlx | 45.26 | 54.60 | +20.6% |
| gemma4:12b-mlx-bf16 | 27.70 | 48.84 | +76.3% |
| qwen3.5:27b | 21.98 | 48.36 | +120.0% |
| gemma3:27b | 21.32 | 47.74 | +123.9% |
| qwen3:32b | 18.52 | 42.88 | +131.5% |
| gemma4:31b | 18.72 | 39.75 | +112.3% |
| qwen3.8:27b-mlx-bf16 | 18.06 | 36.65 | +102.9% |
| gemma4:31b-mlx-bf16 | 12.94 | 26.11 | +101.8% |
| deepseek-r1:70b | 8.82 | 22.65 | +156.8% |
| llama3.3:70b | 8.93 | 22.57 | +152.7% |
| command-a:111b | 5.58 | 15.38 | +175.6% |
| mistral-large:123b | 5.20 | 13.69 | +163.3% |
Generation speed in tokens per second, average of runs 2–11 in Ollama, 2,000 output tokens, 8k context. The faster machine is in bold. Source: Vše Smart.
Round one: small models
Even with small models the Mac Studio wins clearly: Qwen3 4B runs at 203 tokens per second against 121, gpt-oss 20B at 169 against 101. But both machines are fast enough here for comfortable work. The only round the MacBook took was Gemma 4 E4B, and only just: 93.12 against 92.95 tokens per second, effectively a draw.
Round two: where twice the hardware is not twice the speed
Doubling the cores and memory did not double the speed everywhere. Qwen 3.5 35B in MLX gained only 17 percent, Qwen3-Next 80B 29 percent. Two of the smallest gains came from mixture-of-experts models that use only about 3 billion parameters per token. At the other end, Qwen 3.8 27B in MLX was 139 percent faster on the Studio.
Round three: the heavyweights
This is where the Mac Studio delivers the knockout. With Llama 3.3 70B it writes 22.6 tokens per second against 8.9 on the MacBook, with DeepSeek R1 70B 22.7 against 8.8. Command A 111B and Mistral Large 123B are 2.6 to 2.8 times faster. Under ten tokens per second, waiting for an answer gets tedious; above twenty, a 70B model becomes a tool for everyday work.
Verdict
The Mac Studio M5 Ultra wins 31 to 1, with a median advantage of about 89 percent. According to the article it was also the quieter of the two. The MacBook Pro M5 Max remains an impressive machine for small and mixture-of-experts models, and it fits in a bag. If you mainly run models of 70 billion parameters and more, the Studio is in another league, and its 256 GB also takes models that do not fit into 128 GB at all, such as Qwen3 235B.
As always, the test measures speed, not the quality of answers, with one type of prompt in Ollama. More results from the same measurement are in our article on 15 models on the Mac Studio M5 Ultra.
Try the winner
You don’t have to buy the champion to use it. Rent the Mac Studio M5 Ultra with 256 GB from one day and run your own models on it.