How we rank free AI models

FreeLLMRank shows the best AI models you can use without paying, ranked by real human votes, with what you need to run them.

The ranking

Scores come from Arena's public leaderboard dataset. On Arena, people send a prompt, get two anonymous answers and vote for the better one; the votes become a score with the Bradley-Terry model. We keep only the models you can use for free: open weights, or a free API today. Rank numbers (“#11 of 416”) count all models, paid ones included.

Sizes and memory

Parameter counts come from the official Hugging Face repositories (safetensors metadata). Context memory comes from each model's configuration file: layers, attention heads, sliding windows and compressed attention. For repositories that require accepting a license, we read the same configuration from Unsloth's public copy.

Speed

Speed is an estimate: memory bandwidth × 70% ÷ the data read for each token (active weights plus context). It matches real llama.cpp results within about ±30%. Graphics card specifications come from the manufacturers' data sheets.

Small models without votes

Arena votes mostly on large models. We also list popular small models (0.5 to 16 billion parameters) from the labs' official Hugging Face pages, marked “No Arena votes”. They are ordered by downloads or size, never presented as ranked by quality.

Data sources

Arena leaderboard dataset, CC BY 4.0, by Arena Intelligence. We filter it and add sizes; we are not affiliated with Arena.
OpenRouter model list, to find models with a free API.
Hugging Face Hub API, for parameter counts, configurations and downloads.

Updates

Everything updates every night: rankings, free APIs, sizes, model pages and the weekly report on the blog. Last update: 18 Oct 2026. The code is open source.

Frequently asked questions

What about brand-new models?

Arena takes a few days or weeks to collect enough votes. Meanwhile, new models from the labs’ official pages appear in “New, waiting for Arena votes”, with their size, memory needs and speed. When Arena votes, they move into the ranking automatically.

What counts as a free model?

A model with open weights that you can download and run yourself, or a model you can call through a free API today (we check OpenRouter every night). Paid-only models are left out of the rankings, but we show the best paid score as a reference line.

Where do the scores come from?

From Arena (formerly LMSYS Chatbot Arena), where people compare two anonymous answers and vote for the better one. Arena publishes its leaderboard as an open dataset under the CC BY 4.0 license; we filter it to free models. We are not affiliated with Arena.

How often is it updated?

Every night. A scheduled job downloads the latest Arena data, the free model list from OpenRouter and model sizes from Hugging Face, then rebuilds the whole site, including the weekly report on the blog.

Can I run these models on my computer?

Open models, yes, if they fit in your GPU memory. The Run at home page estimates memory and speed for your exact card, from one to ten cards, using each model's real architecture.

Why do small models have no score?

Arena votes mostly on large models. We list the most popular small models from the labs' official Hugging Face pages separately and mark them “No Arena votes”, so you can still find something that runs on a 4 to 12 GB card.