Mac Compute Rentals › Benchmarks
LLM benchmarks on an M3 Ultra 512GB Mac Studio
Measured on our rental machines, the same hardware you get: M3 Ultra, 32-core CPU, 80-core GPU, 512GB unified memory. Batch size 1, weights from the shared model folder.
| Model (MLX) | Weights | Prompt | Prefill tok/s | Generation tok/s | Peak memory |
|---|---|---|---|---|---|
| gpt-oss-120b MXFP4-Q8 | 59 GiB | 2,048 | 1,920 | 87.8 | 64.3 GB |
| gpt-oss-120b MXFP4-Q8 | 59 GiB | 16,384 | 1,670 | 75.7 | 64.9 GB |
| Llama 4 Maverick 17B-128E Instruct 6-bit | 304 GiB | 2,048 | 277–635 (avg 465) | 36.4 (34.4–39.0) | 326.7 GB |
| Llama 4 Maverick 17B-128E Instruct 6-bit | 304 GiB | 16,384 | 552 | 31.3 | 328.6 GB |
Measured Oct 14, 2026 with mlx_lm.benchmark (mlx-lm 0.32.0, mlx 0.32.3, Python 3.12), 256 generated tokens, 3 trials at 2K and 2 at 16K after a warmup run, averages shown. GPU wired limit 480,000 MB. Maverick's 2K prefill varied between trials because of background disk activity on that machine at the time; the 16K runs were steady.
Reproduce it yourself
uv venv ~/bench && uv pip install --python ~/bench/bin/python mlx-lm ~/bench/bin/mlx_lm.benchmark --model /Users/Shared/models/by-name/gpt-oss-120b-MXFP4-Q8 -p 2048 -g 256 -n 3
Coming next
DeepSeek-V3.2 4-bit (352 GiB) once it finishes caching on Studio A. See what else fits on the 512GB fit table.
Run your own model on a whole 512GB Studio
3 days $299 · 7 days $489 · 14 days $779. Bare metal, one renter per machine.
Check availability & book