This is great. It solves two of my issues:

  1. It is difficult to figure out if dropping $10k+ on hardware will make a meaningful difference for running local models.
  2. Downloading each model and each quantization of that model to see the speed seems rude to free services like hugging face, so now I can see things like "with my hardware, going to a Q8 only results in 3 more tokens per second"

My one complaint is picking my hardware (9070XT) and Qwen3.6 27B FP16 results in the message:

> This combination won't run. Qwen 3.6/3.8 27B at FP16 needs 54.6 GB — the AMD RX 9070 XT 16GB has 14.9 GB usable. No quantization of this model fits; pick a smaller model or more memory.

While it is true that it won't fit in your GPU memory, it demonstrably does run on my hardware. Llama.cpp supports putting some layers on the GPU and running others on the CPU.

Also, that message says "No quantization of this model fits" but switching to a different quantization of that model (for example, Q8_0) makes the message go away?