MiniMax H3 Max is a video-generation variant built on H3 and post-trained by fal.ai. MiniMax's official Design page offers text- and image-based creation, with videos up to 15 seconds long. For programmatic integration, fal publishes separate endpoints with their own input schemas.
Access Routes
| Workflow | Entry point |
|---|---|
| Create and refine clips in a creative workspace | MiniMax Design H3 Max |
| Generate from a written scene | fal text-to-video API |
| Guide the opening or ending frame | fal image-to-video API |
| Combine reference assets | fal reference-to-video API |
Choose the route before preparing assets or writing integration code. A workspace's controls, account entitlements and parameter labels should not be assumed to match an API's fields.
H3 Max and H3 Base
MiniMax's open-weight announcement describes H3 Base as the generation module within a larger system: Context-IR processes context, Base generates 768p audio-video, and Regenerate-2K performs the higher-resolution regeneration stage.
H3 Max is a post-trained variant. Downloading Base weights does not reproduce the Max hosted service. For the broader checkpoint and licensing background, see AI Rank's existing H3 open-weights guide.
A First Text-to-Video Request
Start with one subject, one action and a simple camera movement. An original prompt to try is: a white ceramic cup on a wooden table, steam rising slowly, with a gentle forward camera move and steady morning side light. Check the cup's shape and camera movement before adding more objects or actions.
The following input object uses fields documented for fal's text-to-video endpoint. Supply it as the input to minimax/h3-max/text-to-video; authentication and request handling are separate steps.
{
"prompt": "A white ceramic cup on a wooden table. Steam rises slowly. The camera gently moves forward. Soft morning side light. No subtitles.",
"duration": 5,
"resolution": "768P",
"aspect_ratio": "16:9",
"prompt_expansion_mode": "disabled",
"enable_safety_checker": true
}
Keep the original input alongside each result, and change one creative condition at a time. This makes comparisons more useful than changing the prompt, resolution and framing together.
Image and Reference Constraints
For image-to-video, image_url supplies the opening frame and end_image_url can supply the ending frame. Omitting both routes the request to text-to-video. The documented resolution options are 480P, 768P and 1080P; 1080P uses latent refinement from a native 768P source. That setting does not establish native 2K output.
The reference-to-video schema allows at most 12 image, video and audio files combined. Individual video clips must be 2–15 seconds long, with no more than 15 seconds of video in total; reference audio has the same individual and combined duration limits. Refer to assets by modality and order, such as “Image 1” and “Video 1,” so the intended roles are explicit.
Before integration, inspect continuity, subject consistency and the returned clip's actual duration. Check the terms for the service you use separately from the H3 Base weight license; they govern different access routes.
Prepared with automated editorial assistance; official pages and API documentation checked on October 22, 2026. No generation jobs or independent performance tests were run.