MiniMax H3

MiniMax H3 is a general-purpose multimodal generation model from MiniMax. MiniMax positions it as understanding a unified context across text, images, video and audio, and generating picture and sound from that context. The JIUFENG workspace exposes its video generation with three starting points: text alone; a first frame that fixes the opening of the shot, with an optional last frame; or references — images, video and audio the model draws on for subject, motion and sound. Generation runs server-side, and the finished clip appears in your creations.

The three starting points give different degrees of control. With text alone, the model decides both composition and motion. With a first frame supplied, the content of the picture is fixed by that frame and the model generates only the motion that follows. References sit between the two, constraining subject and style without pinning every frame. The audio track is generated together with the picture, so no separate pass is needed. The workspace and the API expose the same capabilities and differ only in defaults: calling the API without a resolution gives you 2K, while the workspace defaults to 768P.

Each clip below demonstrates one of the capabilities described above. Play them to see — and hear — the result.

1. Open the workspace with this model selected. Describe the subject, action and camera movement, then choose text only, first and last frames, or reference media.
2. Choose the output size and duration. First/last frames and reference mode are separate options; check the estimated credits for your selection.
3. Sign in to submit the video. When it is ready, play the result with sound and adjust the prompt or input media for the next attempt.

Run in WorkspaceCompare image and video pricing

1. Output size and duration together decide what a run costs, and both are chosen before you submit. Use a lower size and the shortest duration while settling the direction.
2. State how things move — a push-in, a character turning, water rippling — rather than only what is in the frame.
3. When the picture has to be exact, generate a still first and use it as the first frame.
4. Reference video is billed at the output rate and added to the billed seconds: a 10-second clip on a 5-second output bills as 15 seconds. Shorter references cost less.

Input

A text description (required), plus one of three starting points: text only, a first frame (with an optional last frame), or references (up to 9 images, 3 video clips and 3 audio clips, with reference video totalling 15 seconds or less). Images are 20MB each, as PNG / JPG / WebP, or HEIC (converted in your browser before upload).

Duration

5–15 seconds, any whole number, chosen before you submit.

Output

768P / 2K at 24 fps, with a stereo audio track.

Rate

768P at 0.75 credits/sec, 2K at 1.56 credits/sec. Reference video is billed at the output rate and added to the billed seconds; the first 5 reference images are free and 1 credits each beyond that; reference audio is not charged.

Failure refund

Credits spent on a generation that fails because of a system error are returned automatically.