Apple M3 Ultra for local LLMs

12 measured results from 4 builds, each linked to its source.

Can it run? DeepSeek R1 671B on Apple M3 Ultra Mistral Medium 3.5 128B on Apple M3 Ultra Muse Spark Open 816B-A42B on Apple M3 Ultra Qwen3.8 27B on Apple M3 Ultra gpt-oss 20B on Apple M3 Ultra

All builds with the Apple M3 Ultra

All measured results

Hide speculative-decoding results ⚡ (MTP, DFlash, DSpark and similar draft-and-verify methods give much higher tok/s on predictable text and are not comparable 1:1 with plain decode)

All Apple M3 Ultra builds: see also all GPUs · best LLMs by memory size · download the dataset.

FAQ

How fast is the Apple M3 Ultra for local LLMs?

The fastest plain-decode result we have is 154.6 tok/s on Llama 3.2 3B (, llama.cpp) in Mac Studio M3 Ultra 512GB (Geerling test). Speed depends heavily on the model, quant and engine, so compare like with like in the table. RigBase lists 12 measured results from 4 builds.

How much memory does the Apple M3 Ultra have?

unified memory: 512 GB in our builds.

Which models have been measured on the Apple M3 Ultra?

DeepSeek R1 671B, DeepSeek R1 Distill 14B, Llama 3.1 405B, Llama 3.1 70B, Llama 3.2 3B, Mistral Medium 3.5 128B, Muse Spark Open 816B-A42B, Qwen3.8 27B, gpt-oss 20B.

Where do these numbers come from?

Every row links to a build page with its original public source (benchmark logs, READMEs, forum posts). RigBase does not re-run the benchmarks. Results marked ⚡ use speculative decoding and are not comparable 1:1 with plain decode.