Apple M3 Ultra — Benchmarks & Specs

Bottom line: how fast is the Apple M3 Ultra?

SpecPicks median: 14.1 tokens/sec across 3 cited single-user runs of 65-73B dense models at 4-bit (Q4) under ollama (range 13.7–16.0 tok/s), from LLMCheck, Jeff Geerling ai-benchmarks, Hostbor. For local LLM inference it generates 15.0 tokens/sec running glm5.2:753b-a40b (a mixture-of-experts model with ~40B active parameters, so dense models of similar size run far slower) at q4_K_M under mlx (the source did not report the machine's memory configuration), per LLMCheck. Unified memory is the binding constraint for local inference: the Apple M3 Ultra ships in configurations of up to 512 GB, per the manufacturer spec sheet, and a 753B model at q4_K_M needs roughly 457 GB for its weights alone (753B parameters at ~4.85 bits per weight, per llama.cpp), so it runs only on configurations with more unified memory than that. In Geekbench 6 Metal (GPU) it scores 259,668 points, per MacRumors. Across all its cores it scores 46,243 points in Cinebench R23 multi, per GadgetVersus.

Every figure above is a row in the tables below, and each row links out to the review or public benchmark database the number was taken from; the memory ceiling and the weight-size estimate are not table rows and link to their sources inline; the SpecPicks median is computed here from the runs it names, each a row in the AI table. SpecPicks aggregates published measurements; it does not report first-party benchmark runs.