Sunk Cost sunkcost.ai Data checked 2026-09-03

Can a Mac Studio M3 Ultra, 96GB run local LLMs?

Yes — 29 of the 39 open models on this site fit in its 72 GB of usable memory, the strongest being Qwen3.8 27B. Whether that saves you money is a different question, and the answer is usually no.

Price$3,999 at launch — discontinued
Memory96 GB, about 72 GB of it addressable by the GPU at 819 GB/s
Best model it runsQwen3.8 27B — Sonnet-class, 33 tok/s
Pay-back against the APIPays back in 51 years at 500k tokens a day

Run the numbers on this machine

What it runs

ModelSpeedClassGood atMemory
Qwen3.8 27BQ4_K_M 33 tok/s Sonnet-class 19 GB
Qwen3.6 27BQ4_K_M 32 tok/s Haiku-class 19 GB
Qwen3.6 35B-A3BQ4_K_M 98 tok/s Haiku-class 23 GB
Muse Glimmer 30BQ4_K_M 34 tok/s Haiku-class 18 GB
Gemma 4 26B-A4BQ4_K_M 72 tok/s Haiku-class 18 GB
Gemma 4 31B itQ4_K_M 24 tok/s Haiku-class 26 GB
Granite 4.2 30BQ4_K_M 23 tok/s Haiku-class 26 GB
GLM-4.7-FlashQ4_K_M 70 tok/s Haiku-class 20 GB
Gemma 4 12BQ4_K_M 77 tok/s Below every hosted tier 8.0 GB
Qwen3.5 9BQ4_K_M 91 tok/s Below every hosted tier 6.8 GB
Nemotron 3.5 Lightning 30B-A3BQ4_K_M 94 tok/s Below every hosted tier 26 GB
MiniCPM5 2BQ4_K_M 207 tok/s Below every hosted tier 3.0 GB

17 more fit; the calculator lists them all.

The specifics

Chip
M3 Ultra — 28-core CPU / 60-core GPU
Memory bandwidth
819 GB/s
Usable by the GPU
72 GB — macOS lets the GPU wire roughly 75% of unified memory by default (llama.cpp discussion #2182: 66.7% at 32 GiB or less, 75% above). Raise it with `sudo sysctl iogpu.wired_limit_mb`. Memory is treated as GB throughout, which is slightly conservative.
Power under load
270 W (published) — Apple's published 'max' figure for this chip (support article 102027), a CPU+GPU stress number; inference typically draws less. Apple measured the 32-core/80-GPU, 512GB config.
Availability
Discontinued 2026-08-25. Launch price; Apple no longer sells it new, so treat as a refurbished/used reference.
Sources
source 1, source 2, source 3