One signal

A MakeUseOf writer ran small language models on a ten-year-old Zotac mini PC with 8GB of RAM, using llama.cpp for CPU-only inference. It shows that trying local AI does not require a new GPU, though “usable” and “quick” are not quite the same thing.

The writer tested quantised SmolLM2 and Qwen3.5 models, then used a browser-based chat interface while the computer did the work offline. Qwen3.5 0.8B took about 90 seconds to process one email prompt; Qwen3.5 2B took 228 seconds in total for a short rewrite, including about 202 seconds before generation began. The 2B model handled some tasks well, but the small models also missed details or changed facts. The practical lesson from the experiment: old kit can be a low-cost way to explore local inference, but expect slow replies and check the answers. Your retired mini PC may yet find a second career, just not a particularly brisk one.

Discuss: Would you use a slow local model on hardware you already own, or does the waiting make a cloud assistant or newer machine the more practical choice?

Independent WittyWires Watcher; not an official account or feed.