AI models on M3 Ultra, 256 GB

A Mac with an M3 Ultra and 256 GB of memory runs 39 of these 44 models well. The largest chat model that runs well is Qwen3.5 397B. The best-rated model that runs well here is GLM 5.3 Flash at UD-IQ4_XS, which writes about 20 words a second 4K tokens into a chat. Every number here is an estimate from measured llama.cpp tables, not yet a measurement from this Mac.

Sold as Mac Studio (2025), with a 60-core or 80-core GPU.

Fit is the app's own verdict: the memory a size needs at each context length, measured by llama.cpp's calculator when the catalog was built, against this Mac's 256 GB and the share of it the graphics side may use. Runs great and Runs well leave room for the rest of the Mac. Tight fit means quit other large apps first.

Speed is predicted from the chip's memory bandwidth (800 GB/s here), the weights a token reads, a fixed cost per layer, and the cost of attention over the chat so far. The constants were fitted on October 11, 2026 to 108 measured llama-bench rows across every Apple chip, and land within about 5% of those rows. Two independent reports checked in October 2026 ran 14% and 18% faster at an empty chat, so the estimate leans low. A mixture-of-experts model and any chip newer than M5 are extrapolated.

Outmake times every model you run. With Send Measurements and Errors on, your Mac's numbers join a public table and replace these estimates for everyone with the same Mac. The method, the rules and the data.