How Outmake judges a model

Whether it runs on your Mac

Outmake reads the header of the model's own file for its layers, its context and how much memory each word of context takes. From those it works out the memory the model needs at each size, and compares that with what your Mac can give it.

Speed is either measured or estimated, and the row says which. Measured speeds come from Macs running Outmake whose owners switched that on. Estimates come from a formula fitted to those measurements. How both are made, and how far off the estimates run, is on Measurements.

Stars

A model's stars are where it stands, from one to five, across the boards that tested it. LM Arena counts twice, because people judged the answers without knowing which model wrote them. Every other board counts once. Five stars is the best model on a board that Outmake can run, and every other model is marked by its chance of beating that one in a blind vote: a coin flip is five, one win in four is about three. Pictures, editing, video and video from a picture are each rated on their own board. A model that beats the best one here but won't run on Outmake also shows five.

104 models have an LM Arena rating and 85 have a score on another board. A model with no Arena rating needs two boards before it gets stars. Everything else shows no stars. A model on Outmake's own list that nobody has tested yet shows an estimate, marked as one.

Some scores on Hugging Face are filed by the model's own maker. Re-tunes and merges are left out of the stars for that reason: anyone can file a number on their own page.

Sources

Hugging Face

Every chat model's files, sizes, licence, downloads and likesRead each morning

LM Arena

Ratings from people's blind votes, for text, coding, vision, pictures and videoCC BY 4.0 · October 28, 2026

Aider polyglot

Coding exercises solved in six languagesOctober 3, 2025

LiveBench

Reasoning, coding, maths, data and language on fresh questionsJune 25, 2026

MMLU-Pro

Graduate-level questions in 14 subjectsNovember 5, 2026

GPQA Diamond

Questions PhD scientists write for each otherNovember 5, 2026

Humanity's Last Exam

Questions written to be beyond today's modelsNovember 5, 2026

SWE-bench Verified

Real GitHub issues fixed so the project's tests passNovember 5, 2026

Terminal-Bench 2.0

Jobs finished alone in a terminalNovember 5, 2026

AIME 2026

This year's AIME competition problemsNovember 5, 2026

MMMU-Pro

College-level questions about imagesNovember 5, 2026

Open ASR Leaderboard

How many words a dictation model gets wrongRead when you open Voices

Civitai

Community pictures, styles and models, shown from Civitai and never copiedRead as you browse

Piper

Voices and their samplesRead when you open Voices

Outmake's own Macs and users

Speeds measured on real Macs, sent only by people who turned that onEach night

Left out

Artificial Analysis publishes good scores, and its terms do not let anyone republish them or build on them, so they are not used. SWE-bench's own site data is licensed for non-commercial use, so the store reads the scores makers file on Hugging Face in its place. The one public voice arena rates only two open voices, so voices have samples to play and no stars.

Pictures and videos from Civitai are only ever the ones Civitai rates safe for everyone. They load from Civitai when you look at them and belong to the people who made them.

The Top 10

One list for everyone, by a fixed rule, with each week's list kept for good. The rule and every week are on Outmake Top 10.

Search engines

217 model pages are offered to search engines: the models on Outmake's list, and makers' own releases with 50,000 downloads or out in the last 30 days. The rest stay in the store and out of search, because their pages say little that a better-known model's page doesn't.

Corrections

If a number here is wrong, or your model is listed in a way you don't want, write to hello@outmake.app. Chat data was last read on November 5, 2026 and the other tests on November 5, 2026. Outmake is made by Ambedo Labs LLC in Utah.