A 10-year-old mini PC can run AI, but it takes patience

You don't need expensive hardware to run artificial intelligence at home. That is the conclusion of a test from MakeUseOf, in which writer Gregory Gibson tried running small language models on a 10-year-old mini PC. The article was published on October 24, 2026.

Old machine, small models

The machine was a Zotac ZBOX C1325 nano with a quad-core Celeron N3160 processor, 8 GB of memory and a cheap SSD. It ran Ubuntu Server and used only the processor, with no help from a graphics card. Gibson set up the llama.cpp program and used a chat page in the browser.

He tested three small models: SmolLM2 360M, Qwen3.5 0.8B and Qwen3.5 2B. The numbers indicate the model's size. The smaller the number, the less space the model takes up, but typically the less it can do, too. Memory usage ranged from about 390 MB for SmolLM2 to about 1.3 GB for Qwen3.5 2B, so the machine avoided using swap, which is hard drive space used as emergency memory.

Fast or good

Qwen3.5 0.8B took about 90 seconds to process the task and wrote 2.63 tokens per second. A token is a small piece of text, often part of a word. The larger Qwen3.5 2B took about 202 seconds to process the task and wrote 1.23 tokens per second. The speed for SmolLM2 was not reported.

Quality followed size. SmolLM2 performed worst: when it rewrote an email, it left out a deadline and details about a presentation. It extracted the number 12 as the item instead of the quantity, and in a summary about Wi-Fi it reversed facts.

Qwen3.5 0.8B preserved the email's content and extracted data correctly, but did not meet the requirement to write a summary. Qwen3.5 2B handled the email, extraction and summary tasks and solved quadratic equations correctly with verification.

Impractical for everyday use

The wait was mostly because the machine had to process what it was told. Gibson himself estimates that he could have finished the tasks before the answer arrived. According to the article, the experiment shows that local AI can run without expensive hardware, but that the setup is impractical for regular use.

For comparison

A separate test from It's FOSS on an older i5 laptop with 12 GB of memory and Ollama gave a similar picture. Models with 7-9 billion parameters reached 3-4 tokens per second, while models with 0.6-1 billion reached 18-36. It's FOSS considers models with 1-2 billion parameters the best balance on machines with about 8 GB of memory. Responses at 3-5 tokens per second work technically, but according to them feel painfully slow.

It is a different machine and different models, so the numbers cannot be compared directly. Gibson's results are his own and have not been independently tested.