VZMORE’s AX9 Max mini PC ran a 27-billion-parameter Qwen model at about 16 tokens per second in Yanko Design’s test, putting local AI into a compact desktop without a separate graphics card. The trade-off is a $1,499 price, shared system memory and a fan you will hear under heavy workloads.
Watch Desk analysis
What happened
Yanko Design reviewer JC Torres tested the AX9 Max with an AMD Ryzen AI 9 HX 470 processor, 32GB of RAM and a 1TB SSD. The review reports that a Qwen model with 27 billion parameters, quantised to Q4, produced about 16 tokens per second with up to 24GB of system memory allocated to graphics. A smaller, nine-billion-parameter Qwen model reached about 22 tokens per second in a different setup, with 16GB allocated and a 32K-token context window.
The review also describes sustained heavy workloads without observed slowdown, but notes that the cooling fan is clearly audible. Its tested high-resolution monitors worked over USB-C DisplayPort; HDMI flickered before the image disappeared in those tests, though the review did not establish a cause. The 32GB RAM, 1TB SSD configuration is listed at $1,499. Read the Yanko Design review.
Why it matters
Local-model performance is easier to judge when there is a model, a configuration and an actual speed attached. These results suggest the AX9 Max can make experimenting with a sizeable quantised language model practical on a compact desktop. They are not a promise that every model, application or memory setting will run at the same pace.
The memory trade-off matters: graphics allocation comes out of the computer’s system memory, so giving a model more room leaves less for other applications. Buyers also have to weigh the premium price against the appeal of keeping inference on their own machine, rather than relying on a cloud service.
Our read
The interesting result is not that a mini PC can run an AI model. It is that the review gives readers a usable reference point: around 16 tokens per second for a 27-billion-parameter Q4 model in one stated setup. That makes this more than a tidy silver box with “AI” on the label, although one reviewer’s figures are a starting point, not a universal benchmark.
If local inference is the reason you are considering a compact desktop, compare results on the models and workloads you actually use, and leave enough memory for the rest of your computer. If you mainly want a quiet, inexpensive office machine, this one has some explaining to do.
What to watch
- Whether other testers reproduce the local-model results with comparable settings.
- How the AX9 Max handles longer inference sessions and different models.
- Whether a future firmware or driver update changes the reported monitor behaviour over HDMI.
Discussion spark: Would you pay a premium for a compact PC that runs local AI models, or are cloud access and a cheaper everyday desktop the better trade?
Sources and evidence
- Source update (15 October 2026, 15:20 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.