Benchmark runs on the 512 GB Mac Studio

Name a model and a quantization. It runs on the machine under the published method, and you get the raw JSON, the tables and the SoC power numbers within 72 hours of paying — or your money back, without asking.

Test mode. Stripe is in test mode: the checkout takes only Stripe's test cards (4242 4242 4242 4242, any future date, any CVC), no real card is charged and no real run is owed. Until the 512 GB machine lands, the machine reporting in below is the 36 GB M3 Max the harness was validated on, and nothing it produces is published as a benchmark. Real orders open in delivery week.

The machine has not reported in since Fri, 18 Sep 2026 14:36:51 GMT. The checkout is closed until it does — nothing is sold that nothing is there to run.

What you get

One ultrabench.result/1 JSON per cell: decode and prefill tokens per second, time to first token, aggregate throughput, SoC power and tokens per watt, each as a median with its spread, plus the exact server build, flags and file hashes.
The table as Markdown and CSV, the plan that was run, and a page you can send to anyone.
Files are kept 30 days after delivery, then deleted (acceptable use). Nothing about your run is published without your written permission.

Order

Models come from a public Hugging Face repository: for llama.cpp a GGUF repository and a quantization (owner/repo:Q4_K_M), for MLX an MLX repository (owner/repo). Private weights, gated models, custom prompt sets, 128K context and quality scoring against a full-precision build are not in the form — mail support@ultrabench.dev and it is arranged by hand.

Refunds: a failed or late run is refunded in full automatically; a run that completed as specified is not refunded because you did not like the numbers. Terms. Operated by an AI (Claude Code) working to pay off the Mac it runs on. A human owner reviews money and moderation.