Videos · Resources
Safety and limits
What is checked before a job runs, what is not, and the limits that apply.
Moderation
- Prompts are checked before the job is queued. A rejected prompt returns
content_policy_violationand uses no GPU time. - Image, video, and audio inputs for
fl2vaandref2vaare higher risk, because they can be used for likeness manipulation. They are checked for unsafe content and for faces or likeness before the job is queued. These tasks are not to be enabled without that check. - Outputs are not scanned before delivery. That is a known gap. Depending on how you use the service you may need your own output review.
Limits
| Limit | Value |
|---|---|
| Clip length | 4 to 15 seconds per job. |
| Outputs per job | 1 to 10 (num_outputs_per_prompt). |
| Resolution | 768p on the short edge. |
| Inputs | URLs we fetch. Size and content-type limits are not final. |
| Rate limiting | Per API key, using a sliding window. Numbers are not final. |
| Result retention | Short, then deleted. Window not final. |
Rate limiting exists so that one key cannot take over the small, expensive pool of GPU workers and to limit the damage from a leaked key. Expect queued time under load.
Pricing and billing
Model licence
MiniMax-H3 is released under the MiniMax H3 Community License Agreement. Terms for reselling generations through a paid API are being confirmed. See the model page.