Which open-weight LLM actually runs on your hardware? BestLLMfor ranks 239 local models by VRAM, speed and licence, exposes everything as a free CC BY 4.0 JSON API (no key), and ships an MCP server so your agent can query it directly. Plus 5 free calculators (VRAM, local-vs-API cost, quantization).
Hi PH, I'm Mohamed, solo maker.
I run local LLMs on a single RTX 5070 Ti and kept asking the same question: which model fits my VRAM, and how fast will it really be? Every answer was scattered across Reddit threads and model cards.
So I built BestLLMfor:
- 239 open-weight models with VRAM per quantization, tokens/s estimates per GPU, licence and context, plus 86 GPUs and Macs.
- A free open-data API (CC BY 4.0, no key, CORS): https://bestllmfor.com/api/model... - build on it, cache it, fork it.
- An MCP server so Claude Desktop or any MCP client can ask "best local LLM for 16 GB VRAM" and get structured data: https://github.com/MGM-FALCON/qu...
- 5 calculators: VRAM, local-vs-API cost, quantization recommender, stack budget, API cost explosion.
What's next: real measured tokens/s from my bench rig on every new release, and community-submitted benchmarks through the API.
Happy to answer anything about local inference, VRAM math, or the data model. Roast welcome.