Local, open-weight AI models in ClawAI

What running a model locally actually changes

A cloud provider on this site runs a model on its own infrastructure and charges per request. Ollama and llama.cpp instead load an open-weight model onto hardware you control — your own machine, or a server you operate — so the request never leaves it. That changes who can see the request, not what the model is capable of; a locally run open-weight model is a different kind of thing from any of the cloud providers listed elsewhere in this cluster, not a drop-in replacement for one.

Ollama and llama.cpp are two different tools

Both are real ClawAI adapters, but they suit different situations — Ollama focuses on ease of pulling and running a model with sensible defaults, and llama.cpp gives more direct control over how a model is run at the cost of more manual setup. The full comparison lives at Ollama vs llama.cpp, linked below, rather than being repeated here.

Choosing which open-weight model to run

This page deliberately names no specific open-weight model, because the field moves faster than a static page can track and a stale recommendation is worse than none. What is local-first AI, linked below, explains open-weight models and the trade-off against cloud providers in more depth than a product page should.

Questions people ask

Does local AI cost anything through ClawAI?

ClawAI does not charge a per-token rate for a locally run model the way it does for a cloud provider, since there is no cloud provider being billed — the cost is the hardware you already run it on. Confirm current plan behaviour on the pricing page.

Which open-weight model should I run?

This page does not recommend one — see what is local-first AI, linked below, for how to think about the choice, since the right model depends on your hardware and task in a way a static page cannot track responsibly.

Is a locally run model as capable as a cloud model?

That depends entirely on the specific open-weight model and your hardware, and this page will not make a blanket claim either way. See how to evaluate AI models, linked below, for how to check for your own workload.