Best open source LLM rankings

What "open weights" means here

A model is listed here when its weights are published and downloadable — you can run it on your own hardware, fine-tune it, and keep your prompts off a third party's servers. That is a narrower claim than "open source": most of these models publish weights without publishing training data or training code, and several carry licences with commercial restrictions. Check the licence on the model card before you build a business on one.

Why the ranking is the same score

Open-weight models are scored on exactly the same LiveBench tasks and the same published weights as proprietary ones — there is no separate, easier scale. That is deliberate: an open-model ranking is only useful if it tells you what you give up, if anything, by self-hosting. Where an open model reaches the top of the overall index, that is a like-for-like result, not a handicap category.

The price column still applies: most open-weight models here are also served by hosted providers, and the figure shown is that hosted price. If you self-host, your real cost is your own compute instead, which is usually cheaper at sustained volume and more expensive at low or spiky volume. Context length, by contrast, is a property of the model and holds wherever you run it.

Models with no score are absent from the current benchmark snapshot. They keep their factual data — price, context, modality — and make no capability claim, rather than being quietly ranked last. See methodology.