M1 Ultra LLM benchmark
No signed M1 Ultra submissions yet. Check the configuration guidance below, or contribute a reproducible measurement from your own machine.
Specs vendor specs, not measured
- Unified memory (max configuration)
- 128 GB
Plan memory before choosing a model
The catalog lists a family maximum of 128 GB of unified memory. Enter your installed amount; smaller configurations have less space. Model weights, context cache and runtime overhead must share the available memory. Capacity alone does not establish which model will fit or how fast it will run. Use the memory-fit calculator for your model and context as an estimate, then verify with the intended runtime. No signed M1 Ultra measurement is available yet.
Contribute a signed M1 Ultra benchmark
Run on YOUR hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Community folklore on M1 Ultra
2 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 60%
12.00tok/s — Qwen2.5-72B on M1 Ultra via lm-studio
our signed data: M1 Ultra · Qwen2.5-72B
“By comparison, my M1 Ultra does about 12 tokens/s for Qwen2.5-72B-Instruct (4bit). The extra bandwidth is just insanely good. BTW One other thing you can try is using speculative decoding”
- communityconfidence 55%
12.70tok/s — on M1 Ultra via mlx fp16
our signed data: M1 Ultra
“nd, vanilla JS frontend, WebSocket streaming **Some benchmarks on M1 Ultra (64GB):** |Model|Speed|Notes| |:-|:-|:-| |GooseOne 2.9B (fp16)|12.7 tok/s|Constant memory, no KV cache| |Z-Image Turbo (Q4)|77s / 1024×1024|Metal acceleration via mflux| The RNN advantage that made me b…”
Request a run on M1 Ultra
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy