EN
26 October 2026 · Jakub M.
Mac Studio M5 Ultra under test: Qwen3 235B at 46 tokens per second and fast prompt reading
The Czech tech site Vše Smart tested the Mac Studio M5 Ultra with 256 GB on five language models in llama.cpp and MLX. The largest, Qwen3 235B, took 133 GB of memory and generated 46 tokens per second. Here is what the results mean if you want to run large models.
Source: Mac Studio M5 Ultra v testu: Qwen3 235B a rychlost čtení promptů, Ludvík Eckerman, Vše Smart, 26 October 2026 (in Czech). All figures below come from that article.
The Czech site Vše Smart published a detailed test of the Mac Studio with the M5 Ultra chip. It is the same configuration we rent out: 80-core GPU and 256 GB of unified memory. The test focuses on two things our own first benchmark did not cover: how fast the Mac reads long prompts, and a model too large for a 128 GB machine.
How Vše Smart tested
- Hardware: Mac Studio M5 Ultra (80-core GPU), 256 GB, macOS 27.0.1, with nothing else running.
- Software: llama.cpp with all layers on the GPU, and MLX 0.32 with mlx-lm. Power and temperatures logged with macmon.
- Models: Qwen3 8B, Qwen3 30B-A3B, Llama 3.3 70B, gpt-oss-120b and Qwen3 235B-A22B, all in 4-bit (gpt-oss in MXFP4). Qwen3 235B in MLX only.
- Tests: prompts of 512, 4,096 and 16,384 tokens with 128 generated tokens, plus ten minutes of continuous load. Measured on 24 October 2026.
Results
| Model | Memory | Generation, llama.cpp | Generation, MLX | Prompt 4k, llama.cpp | Prompt 4k, MLX |
|---|---|---|---|---|---|
| Qwen3 8B | ≈ 5 GB | 149 | 176 | 4,205 | 6,360 |
| Qwen3 30B-A3B | ≈ 18 GB | 172 | 159 | 4,535 | 9,870 |
| Llama 3.3 70B | ≈ 42 GB | 23 | 25 | 509 | 716 |
| gpt-oss-120b | ≈ 63 GB | 142 | 130 | 3,387 | 6,538 |
| Qwen3 235B-A22B | ≈ 133 GB | – | 46 | – | 1,656 |
Tokens per second with an empty context: generating 128 tokens, and reading a prompt of 4,096 tokens. Qwen3 235B was tested in MLX only. Source: Vše Smart, measured on 24 October 2026.
MLX reads prompts up to twice as fast
Before a model writes its first word, it must read the whole prompt, for example a document you want summarised. According to Vše Smart, MLX read prompts 1.4 to 2.2 times faster than llama.cpp with every model; the gap was largest with Qwen3 30B-A3B. With gpt-oss-120b in MLX, a 4,000-token document is read in well under a second. For generation neither framework was a clear winner, so it is worth trying both with your own model.
A long context
| Model | Generation | Reading the prompt |
|---|---|---|
| Qwen3 8B | 121 | 5,240 |
| Qwen3 30B-A3B | 115 | 6,822 |
| Llama 3.3 70B | 22 | 650 |
| gpt-oss-120b | 114 | 6,332 |
| Qwen3 235B-A22B | 33 | 1,289 |
MLX, prompt of 16,384 tokens followed by 128 generated tokens, tokens per second. Source: Vše Smart.
With a 16,384-token prompt, generation in MLX slowed by about 12 to 31 percent, depending on the model.
Qwen3 235B needs more than 128 GB
The largest model took about 133 GB of memory and generated about 46 tokens per second, faster than most people read. It would not fit on a Mac with 128 GB, which is where the 256 GB configuration makes the difference.
Temperatures, power and the M3 Ultra
The GPU peaked at 87 °C and the chip drew up to 215 W while reading prompts. The fan rose to 2,265 rpm, 62 percent of its maximum, and during ten minutes of continuous load the test saw no throttling. In the standard llama.cpp test with Llama 2 7B (Q4_0), the article estimates the M5 Ultra reads prompts about 3.5 times faster and generates about twice as fast as the M3 Ultra. The author points out that the M3 Ultra figures come from the community llama.cpp table measured with older versions, so the comparison is only rough.
As the article notes, the test measures speed, not answer quality, uses only 4-bit quantisation and single requests, and reports chip power rather than whole-system power.
Run it yourself
You can rent a Mac Studio M5 Ultra with 256 GB from one day and test your own models on it. Our one-hour load test shows how it behaves under sustained load.