Which local AI models run on a Mac, how fast they work and how well they're rated. 3,940 chat models, plus picture and video models, on 59 Apple silicon Mac configurations from M1 to M6, refreshed every morning.
October 2026 edition · method v1 · data of November 5, 2026
LM Arena rating, or Outmake's estimate where the Arena hasn't rated a model · higher is better
K2-Horizon 7B is likely the best model that runs great on this Mac, at about 6.1 words a second. LM Arena hasn't rated it yet, so its place is an estimate from its size and age; the best with a measured rating is Gemma 3 12B.
Outmake Bench · Best rated
Best-rated models on a Mac with M4 and 16 GB
LM Arena rating, text with style control · higher is better
Gemma 3 12B is the best-rated model that runs great on this Mac, at about 7.2 words a second. Gemma 3n E4B is next.
Outmake Bench · Rating and speed
Rating against speed on M4 with 16 GB
LM Arena rating against words written a second · up and right is better
5 models sit on the best trade-off line on this Mac, from Gemma 3 12B, the best rated, to Llama 3.2 1B, the fastest. Every model off the line is beaten on both by one on it.
Outmake Bench · Best for coding
Best for coding on M4 with 16 GB
LM Arena coding rating, with style control · higher is better
For code, Granite 4.2 3B leads on this Mac with 1378, ahead of Granite 4.2 8B at 1370.
Outmake Bench · Memory ladder
Memory ladder: the best-rated model each memory size runs
LM Arena rating of the best-rated model that runs great or well · higher is better
With 16 GB the best-rated model that runs well is Gemma 3 12B (1342). The next step up is 24 GB, which runs Gemma 4 26B (1438).
Outmake Bench · Chip ladder
Chip ladder: Qwen3.5 9B on every Apple silicon chip
Words written a second at Q4_K_M (4.94 bits), a few exchanges into a chat · bars show the likely range
An M5 Ultra writes Qwen3.5 9B about 8.9 times as fast as an M4: 85 words a second against 9.5.
Make your own
Chart any two numbers for any Mac, filter, sort and compare, or paste a link to a model and see how it runs.
3,940 chat models people download as GGUF on Hugging Face: the most downloaded, down to about a thousand downloads a month, this month's trending and most liked, and open models a quality board rates. A model counts on a Mac when it runs great or well, the app's own verdict, at 4 bits a weight or more: a rating describes the full model and holds at about that size.
Speed and memory
Estimates for llama.cpp, the engine Outmake runs, from a model fitted to measured runs on each Apple silicon chip. It leans low: two independent reports checked on 11 October 2026 ran 14% and 18% faster at an empty chat. Memory comes from each file's own header and is fitted to the memory the app measures. Method.
Ratings
LM Arena's ratings with style control, overall and for coding, from its leaderboard dataset (CC BY 4.0): 104 models rated. A rating goes only to the maker's own model. Best for this Mac also counts the models on Outmake's list that the Arena hasn't rated yet, at an estimate from their size and release date, drawn with its range and marked as an estimate; no other chart uses an estimate of a rating. Outmake's own data is free to reuse under CC BY 4.0: download it; maker logos are their owners' and not part of it.
Outmake Bench, October 2026 edition (method v1, data of November 5, 2026). https://outmake.app/bench. CC BY 4.0.
