willitrun

Check which AI models your PC can actually run, and how fast. Auto-detects your GPU, VRAM & RAM. 100% offline.

GPU: Intel(R) Arc(TM) 140T GPU (15.7 GB, integrated)   RAM: 31.4 GB

 Model              Quant   Needs     Status        Speed       Run with
 GPT-OSS 20B        Q4      14.5 GB   ✅ Fits       ~28 tok/s   ollama run gpt-oss:20b
 Qwen3 14B          Q4      10.2 GB   ✅ Fits       ~7 tok/s    ollama run qwen3:14b
 GLM 4.7 Flash      Q4      20.7 GB   🐢 CPU only   ~20 tok/s   ollama run glm-4.7-flash
 Llama 3.3 70B      Q4      48.7 GB   ❌ Too big    -           ollama run llama3.3:70b

Best for you: Qwen3 30B A3B Q4 (~18 tok/s) → ollama run qwen3:30b

Install & run

Requires Python 3.9+.

git clone https://github.com/kenttsimon/willitrun
cd willitrun
python -m pip install -r requirements.txt
python willitrun.py

How it works

1. Detects your hardware. NVIDIA via nvidia-smi, Intel/AMD on Windows, and Apple Silicon.
2. Checks if each model fits. Size ≈ parameters × bits per weight, plus about 15% for context and runtime. It tries Q8 first, then Q4.
3. Estimates speed. Text generation is limited by memory bandwidth, so tokens/sec ≈ bandwidth ÷ model size.
4. Handles MoE models. Mixture-of-experts models only read their active parameters per token, so a 30B model with 3B active runs about as fast as a 3B model.

Speeds are estimates. Real results depend on drivers, context length and background apps.

Safe to run

• Offline. No internet calls, your specs never leave your machine.
• Read-only. It only reads hardware info and never changes files or settings.
• Minimal dependencies. Only psutil and rich.

Supported hardware

Type	Detection
NVIDIA GPUs	✅ Automatic
Intel / AMD integrated (Windows)	✅ Automatic (uses shared RAM)
Apple Silicon	✅ Automatic (uses unified memory)
AMD / Intel dedicated GPUs	⚠️ Name detected, VRAM from database
Linux AMD / Intel	🚧 Not yet

Contributing

• Add a model: add a line to data/models.json (params_b, plus active_b for MoE models).
• Add a GPU: add its VRAM and memory bandwidth to data/gpus.json.
• Share real benchmarks: open an issue with your GPU, model and actual tokens/sec. It helps make the estimates more accurate.

Run python test_estimate.py before opening a pull request.

License

MIT