Acceptable Use Policy

Effective 8 September 2026. Ultrabench · ultrabench.dev

This applies to everything you can send us: models and configurations for benchmark runs, files for batch jobs, mail to our addresses, and messages in our stream chat.

Do not send us

• Anything unlawful, or anything whose processing would be unlawful where we or you are.
• Child sexual abuse material, or non-consensual intimate imagery. Both are reported to the authorities and to the platform involved; no refund, no appeal.
• Malware, exploits, or content built to attack, defraud, impersonate or harass someone.
• Credentials, keys, tokens, or data you do not have the right to hand over.
• Regulated personal data — health records, payment card data, government identifiers, biometrics, children's data. We offer no HIPAA, PCI, FedRAMP or equivalent compliance and cannot lawfully process it. Anonymise or synthesise before you send.
• Model weights you are not licensed to use. By submitting weights you confirm you have that right; we only ship products built on permissively licensed models.
• Bulk personal data for scraping, enrichment, surveillance or profiling of individuals.

Do not use the service to

• Mine cryptocurrency, crack passwords or hashes, or run anything whose point is to consume compute rather than measure it.
• Generate spam, disinformation campaigns, or synthetic reviews and engagement.
• Circumvent our rate limits, quotas or pricing, or resell our compute as raw inference. We do not sell raw inference; we sell measurements and finished work.
• Attack the service or the platforms it runs on, including load testing you did not clear with us first.

How we handle what you send

• Ownership. Your inputs and the results stay yours. We keep the right to publish aggregate, de-identified performance statistics — "quantised 70B models averaged N tokens/second across M runs" — and never your files, prompts, model names or identity without written permission.
• Retention. Inputs and outputs are deleted within 30 days of delivery, sooner on request. Job metadata (timestamps, runtime, cost) is kept for accounting.
• Training. We do not train or fine-tune any model on your submissions.
• Isolation. Jobs run on one machine that accepts no inbound connections; work is pulled from a queue rather than pushed to the Mac. Treat it as a single-tenant box operated by one small business, not as a hardened multi-tenant cloud.
• Turnaround. Each job states its turnaround before you pay. Miss it and you are refunded automatically (refunds).

Stream chat and public spaces

Moderated by an automated classifier plus the platform's own tooling, with a human owner holding the final word. Harassment, slurs, doxxing and spam get removed and the account timed out or banned. Moderation decisions are appealable to support@ultrabench.dev and a human reads the appeal.

Enforcement

We may refuse, pause or cancel any submission at our discretion, and we refund cancelled work — except where the submission was itself illegal. Repeated breaches end service. Where the law requires a report, we make it. If you think we got an enforcement call wrong, mail us; a human reviews it.

Reporting abuse

Mail support@ultrabench.dev with "abuse" in the subject. Anything involving imminent harm reaches the human owner directly, not the automated operator.