A November update for anyone planning a local setup: the catalog now has 160 hardware presets and 350 models.
Recent additions include MiniCPM5 2B, LFM2.5 24B A2B, and LFM3 1.7B. I have also added full-system budget estimates with personal-use defaults, so you can compare the GPU alongside the supporting components.
One concrete compatibility update today: the official LFM3 1.7B Apple Silicon checkpoint now has an MLX-VLM 0.7.11+ recipe. The calculator uses that runtime and the verified 4-bit weights. Other precisions show a verification notice instead of silently switching to 4-bit.
The two frontier charts are still there: VRAM versus published benchmark quality, and GPU reference price versus estimated speed. Memory and speed estimates include their assumptions, and publisher evaluations have source links.
If you try it on a Mac mini, Mac Studio, or a consumer GPU, I would appreciate a reproducible result: hardware and memory, model/checkpoint, quantization, runtime/version, context length, peak memory, and separate prompt-processing and decode speeds. A result that disagrees with the estimate is especially helpful.
Demo: [AI Hardware Fit — 내 GPU에서 돌아가는 AI 모델 찾기](https://jaeseok614.github.io/ai-hardware-fit/?lang=en)
Source: [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](https://github.com/jaeseok614/ai-hardware-fit)
Measurement report: [Sign in to GitHub · GitHub](https://github.com/jaeseok614/ai-hardware-fit/issues/new?template=benchmark-report.yml)
Official LFM3 MLX instructions: [LiquidAI/LFM3-1.7B-MLX-4bit · Hugging Face](https://huggingface.co/LiquidAI/LFM3-1.7B-MLX-4bit)