Will this run on my PC?

Enter your hardware and the model you are considering. The answer is an estimate from parameter count and quantization — the same arithmetic aihwbench fit uses — not a measurement. Where the dataset has a real result for your configuration, it is shown alongside, and it always wins.

Fits entirely in VRAM

An estimated 4.9 GB of weights and overhead fits in your GPU memory, so nothing has to spill to system RAM.

Weights
4.2 GB
With overhead
4.9 GB
Outside VRAM
0%

Where the cliff is

Throughput does not taper as a model outgrows VRAM — it collapses once layers move to the CPU. This is the share that would sit outside GPU memory at each card size.

0 GB32 GB
Share of the model outside VRAM as GPU memory varies. It becomes fully resident at about 4.9 GB — the second line marks your 8 GB. Estimated from parameter count and quantization, not measured throughput.

bits-per-weight are format conventions, not measurements; an unknown quantization is refused rather than guessed. Overhead factor 1.15× covers KV cache, activations and runtime allocation; the KV cache grows with context length, so a long context needs more than this shows.

Measured results at this quantization

  • llamacpp-1789031931-b98435ea — qwen2.5-0.5b-instruct-q4_k_m.gguf on llama.cpp, NVIDIA GeForce RTX 3080 Ti Laptop GPU · 338.3 tok/s · 634.0 MB peak VRAM measured
  • ollama-1789031359-bfb62628 — qwen2.5:0.5b-instruct-q4_K_M on ollama, NVIDIA GeForce RTX 3080 Ti Laptop GPU · 297.1 tok/s · 634.0 MB peak VRAM measured
  • ollama-1789031428-39f82fdf — qwen2.5:0.5b-instruct-q4_K_M on ollama, NVIDIA GeForce RTX 3080 Ti Laptop GPU · 293.2 tok/s · 634.0 MB peak VRAM measured