Models Fit Studio
GitHub
Download
Models Fit Studio scans your hardware, tells you in plain words which local AI model will run well, picks the right file, measures it for real, and serves it to your AI agents.
No settings to learn. Paste a model, read the answer, and go.
Paste any Hugging Face link or name, like unsloth/Qwen3-4B-GGUF. The app reads your CPU, RAM and graphics chip and answers for this PC.
Feel the speed before you spend the download: a sample answer is typed at the pace the model will reach on your machine.
Not sure which one? Put up to four models in a Battle. Each scores points for quality, speed, memory and context on your PC, and one wins.
Open Agents, pick a downloaded model and press Start. Copy the OpenAI or Anthropic address, the key and the model ID into your agent.
Most people find out a model is too big or too slow only after a 10 GB download. This app tells you first.
One verdict per model for your exact hardware, not a table to decode.
Q2, Q4, Q8, F16… the largest version that fits and still writes at reading pace is chosen for you.
Measure speed and test quality on your own PC. Every later estimate is corrected by what was measured.
OpenAI and Anthropic addresses for Claude Code, OpenCode, Codex, Cline, Continue, Aider and more.
Models run on your computer. The llama.cpp engine is built in: nothing else to install.
NVIDIA, AMD and Intel graphics, or the processor alone. NVIDIA cards get their own fast engine by themselves.
Start a model in the Agents tab and point any agent at it. The app writes the exact settings for you, with your key and model filled in.
Get ModelFitStudio-win-x64.zip from the latest release.
Anywhere you like, for example C:\Apps\ModelFitStudio. No installer, no admin rights.
Open ModelFitStudio.exe. If SmartScreen warns, choose More info → Run anyway.
No. It works with NVIDIA, AMD and Intel graphics, and on the processor alone. It reads how much memory your graphics chip can really use, including chips that share system memory.
No. It talks to Hugging Face to list and download models and to GitHub to check for updates. Nothing about you is shared unless you press a Share button yourself.
Before a download they are estimates, shown as a range and corrected for your kind of graphics chip. After Measure speed they are real numbers from your PC, and other estimates are corrected by them.
GGUF models from Hugging Face, the format llama.cpp runs. For a model without GGUF files the app suggests GGUF copies made by others.
Yes. It is open source under the MIT licence: free to use, change and share. No account, no ads, no tracking.