About RunLLMLocal

RunLLMLocal answers one question: can my computer run this AI model, and how well? You pick your graphics card or Mac chip, and it lists popular open-weight models ranked by how comfortably they fit, with the quantization to download and a speed estimate.

Why we built it

Running models locally has become practical for anyone with a gaming PC or a recent Mac, but the first step still trips people up. Model pages list file sizes without explaining whether they fit, and advice online is scattered across forums with conflicting numbers. We kept doing the same back-of-envelope arithmetic for friends, so we wrote it down as a tool.

How the estimates work

Memory needed is the model's weights at the chosen precision, plus 10% runtime overhead, plus a KV cache sized from each model's architecture and your context length. Usable memory is 90% of discrete VRAM or 72% of Apple unified memory. Speed is derived from your hardware's memory bandwidth divided by the bytes read per token. These are the same formulas experienced llama.cpp users apply by hand. They are estimates, and we say so on every result.

Keeping it current

New models arrive monthly. We add notable open-weight releases from Qwen, Meta, Google, Mistral, DeepSeek, Microsoft and OpenAI as they ship, and retire models nobody runs any more. If something you use is missing, tell us on the contact page.

Independence

RunLLMLocal is free and supported by advertising. We do not take payment from hardware or model vendors for placement, and our picks are based on the memory math and public benchmarks only.