16 October 2026 · Jakub Mokříš
Fifteen AI models on the Mac Studio M5 Ultra: how many tokens per second you really get
Czech tech site Vše Smart ran 32 local language models on a Mac Studio with the M5 Ultra and 256 GB of memory. We picked fifteen of them, from the small Qwen3 4B to the 123-billion-parameter Mistral Large, and looked at what the numbers mean for everyday work.
Speed comparisons of local models usually stop after two or three names. The team at Vše Smart measured thirty-two of them on a Mac Studio M5 Ultra with 256 GB of unified memory, each one eleven times. We took fifteen models from their results that cover the whole field well: small and large, dense and mixture-of-experts.
How it was measured
The models ran in Ollama, driven by a simple Python script. Every model got the same task, a long technical text about how modern cars are built, with a limit of 2,000 output tokens, temperature 0 and an 8,192-token context. The first run of each model was thrown away as a warm-up, and the result is the average of the next ten. There was a 30-second pause between models.
The whole test took over four hours, and the models produced more than 700,000 tokens along the way. Only speed was measured; the test did not judge the quality of the answers.
Results
| Model | Type | Generation speed |
|---|---|---|
| Qwen3 4B | dense | |
| Qwen3 30B-A3B | MoE, 3B active | |
| gpt-oss-20b | MoE, 3.6B active | |
| Qwen3.5 35B-A3B | MoE, 3B active | |
| Qwen3.5 9B | dense | |
| gpt-oss-120b | MoE, 5.1B active | |
| Gemma 4 26B-A4B | MoE, 4B active | |
| Qwen3-Next 80B-A3B | MoE, 3B active | |
| Qwen3.5 122B-A10B | MoE, 10B active | |
| Llama 4 Scout | MoE, 17B active | |
| Mistral Small 3.2 24B | dense | |
| Qwen3.5 27B | dense | |
| Gemma 4 31B | dense | |
| Llama 3.3 70B | dense | |
| Mistral Large 123B | dense |
Average of ten runs in Ollama, 2,000 output tokens, 8k context. MoE = mixture of experts; the active figure is the number of parameters used for each token. Data: Vše Smart.
Size tells you little about speed
The clearest example is gpt-oss-120b against Llama 3.3 70B. OpenAI's model has more parameters (117 billion against 70), yet it writes five times faster: 115 tokens per second against 23. The difference is the architecture. gpt-oss-120b is a mixture-of-experts model and uses only about five billion parameters for each token. Llama 3.3 is dense, so it works through all seventy billion every time.
Qwen shows the same thing. The 122B-A10B version, with ten billion active parameters, does 85 tokens per second, while the dense Qwen3.5 27B, a quarter of the size, manages 48. If speed is what you care about, look at the active parameters, not the total.
How fast is fast enough
For ordinary chat, around 20 tokens per second is plenty. At that pace the text appears faster than most people read it. Of the fifteen models in the table, only Mistral Large falls below that line, at 14 tokens per second. It is still useful for jobs nobody sits and waits for, such as processing documents overnight.
Speed matters much more when a coding assistant writes long blocks of code. There, models above 80 tokens per second make sense, for example gpt-oss-120b, Qwen3-Next 80B or Qwen3.5 122B.
It matches our own measurements
We measured several of the same models on our own Mac Studio, using MLX instead of Ollama and a slightly different method. The numbers differ by a few to a few dozen percent, but the overall picture holds. gpt-oss-120b gave us 120 tokens per second (Vše Smart: 115), gpt-oss-20b 184 (169) and Qwen3.5 27B 55 (48). Gemma 4 31B came out slower for us, 31 against 40, because we used a different version of the model. The details are in our benchmark article.
Try it yourself
All thirty-two models fitted into 256 GB of memory with room to spare. If you are thinking about buying a machine like this, or only need the power for a few days, you can rent our Mac Studio M5 Ultra with 256 GB. Daily and weekly rentals are currently free.
Data source: Vše Smart: Mac Studio M5 Ultra tested with 32 local AI models (Ludvík Eckerman, 5 October 2026, in Czech).