Test model speed before buying hardware

Llmspeed gives you a way to get a feel for how fast a local language model might run on different hardware before you buy. The idea speaks to a real purchasing question: whether spending more on a machine will make a useful difference. The Show HN post drew 213 points, and one commenter said the tool helped compare how much faster a higher model quantization might be on their hardware. You can try it through the Hacker News post.

Treat the results as a starting point, not a promise. Commenters said some estimates looked off compared with their own machines, and pointed out that the model list and hardware configurations are limited. Real workloads can also behave differently: one commenter said long prompts and agentic coding tasks can reduce throughput, while others wanted more detail on prefill speed, context size and parallel use. For a business, that means a simulator can help narrow down options, but it won’t tell you exactly how a local setup will perform on your own work. This is the kind of thing we build into an AI command centre, where teams can keep practical AI choices and their limits in view.