AI news · October 18, 2026
Llmspeed lets buyers feel local model speed before hardware spend
Llmspeed compares model speed across 30 GPUs, Macs, and CPUs; use it to narrow hardware choices, then verify the winning setup with a real benchmark.
- listed GPUs, Macs, and CPUs
- 30
- listed models
- 15
- possible estimate movement from setup
- 10-30%
Llmspeed reached a fresh Hacker News discussion on October 17 and was also shared in r/LocalLLaMA. The open tool lets a reader choose a GPU or Mac and a model, then watch text stream at the estimated speed, including the wait before the first token. It has a race mode for two setups and a memory-fit view that shows whether a model fits, needs partial offload, or will not run.
The repository lists 30 hardware choices and 15 popular models. It labels values as measured, derived, or projected and says drivers, runtimes, power, thermals, and attention settings can move real numbers by 10% to 30%. That honesty is the point: the tool is a buying screen, not a benchmark replacement. Use the result to choose two machines, then run the exact model and workload before spending money.
What you can do with it
Pick your likely model and two candidate machines. Compare first-token wait, stream speed, memory fit, and expected prompt length in Llmspeed, then run a real benchmark with the same quantization and context. Record the difference and use it to improve your next hardware estimate.
Our take
This is a useful decision tool because raw tokens per second is hard to imagine before a purchase. The author clearly marks estimates, which makes the tool more trustworthy than a single headline number. It still cannot model every agent workload, so the final check must use your own prompt and runtime.
Links Try itRead the code
Source: AI Weekly Malaysia / Hacker News — Made With Models writes the brief; the reporting is theirs.