Llmspeed estimates how fast a local model will feel on about 30 devices before you buy one

What the tool does

Llmspeed lets you choose a device or a model and see how fast it will "feel". Its own caveat is that figures are "estimates for a short prompt on a single device", partly from community benchmarks and partly projected from specs. The page does not name the metric or unit.

Owners checked it against their own rigs

Owners of the hardware tested the numbers and found them off in both directions:

"I don't know what basis these numbers are estimated on, but my M2 Ultra 128GB runs almost twice as fast as what's reported for the M3 Ultra 512GB." — pil0u, who owns an M2 Ultra 128GB

"It says 46 tok/sec for Qwen3.6-27b 4 bit on RTX5090, but I get 60-70 tok/sec on a small context like that (5 bit weights with 8 bit KV cache)… If I add ngram speculation, then it can increase to 100+ tok/sec on code." — kgeist, who runs that model on an RTX 5090

A commenter who tunes vLLM at work said agentic coding with 60k-140k prompts gives "half to a third" of benchmark decode throughput.

Treat Llmspeed as a floor-to-ceiling sanity check for short prompts. Before spending on hardware, check the cell against a measured result on the same card, and remember long agentic contexts will run slower.