Make your own chart

How the numbers are made

The models. The chat models people download as GGUF on Hugging Face this month (the most downloaded, trending and most liked), refreshed every morning. A compressed copy of a model, such as ISTA-DASLab's GSQ-RCO builds of Qwen3.8, is listed as one of that model's sizes. A fine-tune or a retrained model, such as PrismML's Bonsai, is a model of its own and never borrows another model's score. Updated November 5, 2026; the full list as JSON.

Fit and memory. The size's file, a key-value cache at 8 bits per value for the chat length shown, and a fixed allowance fitted to the memory the app measures on its own models. The verdicts are the app's: runs great leaves room for everything else, runs well leaves enough, a tight fit closes other apps. Every row picks the most bits that runs great, else runs well, from sizes of 4 bits a weight or more first; past about 8 bits a copy only writes slower, so the smaller file wins. Longest chat is the longest conversation that keeps the verdict the model has at 8,000 tokens.

Speed. Predicted, a few exchanges into a chat (4,096 tokens), from the same model the app uses, fitted to measured llama.cpp tables on every Apple silicon chip (constants of October 11, 2026). Rows marked rough are outside those tables: a mixture of experts, a publisher's own size format, or a chip newer than M5. It leans low: two independent reports checked on 11 October 2026 ran 14% and 18% faster than this estimate at an empty chat (Gemma 3 12B on an M1 Max, Gemma 4 26B A4B on an M4). How speed is measured.

Ratings. LM Arena's ratings with style control, overall and for coding, from its leaderboard dataset (October 28, 2026, CC BY 4.0). 104 of the 3,940 models have one; a rating goes only to the maker's own model, never to a fine-tune or a copy under another maker's name. A rating describes the full model, so a size under 4 bits a weight does worse than its rating says; that is why rows lead with 4 bits or more.

Your own model. Paste a Hugging Face link in the search box. Your browser reads the repo's file list and the first few megabytes of one file's header from Hugging Face, and judges it with the same arithmetic. The link is not added to the page's address, and Outmake's servers receive no request for it.