Hi everyone — I built AI Hardware Fit because answering “Will this model run on my GPU?” meant piecing together model files, quantization, context length, VRAM requirements, runtime support, and speed data from different places.
There are already useful tools for estimating whether a model fits in memory. I wanted to connect that question to the next one: which GPU and runtime configuration should I try?
AI Hardware Fit is a free, open-source browser tool that lets you:
- Start with a GPU and see model candidates, suggested quantization, estimated VRAM and speed ranges, and Ollama or llama.cpp commands.
- Start with a model and compare suitable GPUs by budget, power, and form factor.
- Explore a VRAM–quality frontier using scores from the same published benchmark, and a price–speed frontier for compatible GPUs.
The current catalog has 151 GPU presets and 332 AI models. The quick picks include Mac mini M4 (32GB) and M4 Pro (64GB). Since Apple Silicon uses unified memory, the tool shows a GPU memory budget as a calculation basis; the actual amount available to a model depends on macOS and the runtime.
I try to keep the evidence behind the numbers visible. Exact-condition measurements, measurement-calibrated ranges, related measurements, and calculated estimates are labeled separately. Calculated speed is a planning estimate, not a benchmark result or guarantee. The VRAM chart’s baseline is Q4_K_M, 4K context, one concurrent request, and llama.cpp. GPU prices are reference snapshots, not live quotes.
Demo: AI Hardware Fit — 내 GPU에서 돌아가는 AI 모델 찾기
Source (MIT): GitHub - jaeseok614/ai-hardware-fit: Can my GPU run this? Pick a GPU to get 3 fitting AI models, quantization, VRAM, speed ranges, and run commands. 한국어·English. · GitHub
Share a measurement: Sign in to GitHub · GitHub
I’d especially appreciate feedback from people who have tried models locally. Which GPU, model, quantization, runtime/version, context length, and tokens/s did you get? I’m collecting reproducible measurements to improve the estimates and understand where they differ from real hardware.