MiniMax H3 turns a description or a single starting image into a clip of 4 to 15 seconds that carries a generated sound track.
H3 is built around clips that arrive with sound: every render includes an audio track, so a first draft already plays like a finished short rather than a silent rough cut.
Length runs from 4 to 15 seconds in whole seconds, with two quality tiers. That makes it a straightforward choice for short feed and story formats where the audio matters as much as the picture.
In text mode you choose the shape and quality tier. In image mode the shape follows the uploaded still, so the shape selector is replaced by the image upload.
Each render carries a generated audio track, so the first version of a clip already plays as a complete short.
Describe the shot in text, or upload one still as the first frame and let the model build motion from it.
Pick any length from 4 to 15 seconds, which keeps a clip inside common short-form placements.
Vertical or square clips where generated ambience sells the moment without a separate audio pass.
Turn one storyboard line into a playable shot with sound before committing to a longer production.
Animate a hero still and let the generated track carry the reveal instead of adding music later.
Use a button on this page. It opens the right workspace with this model already selected, in the language you are reading now.
Write what should happen in the shot or image. Concrete subjects, framing and action give the model more to work with than short labels.
Choose length or quality tier, shape and sound where offered, then read the credit quote for those exact settings before you generate.
Generation uses credits from your account balance. Video models are charged per generated second at the quality tier you choose, and image models per image. The exact amount is quoted in the workspace before you confirm, and the pricing page lists the current packages.
See credit packagesYes. Every render includes a sound track and there is no silent option for this model. Choose a different video model if you need silent output.
Between 4 and 15 seconds, in whole seconds, chosen in the workspace before you confirm.
768P and 2K. The credit quote changes with the tier and with the length you set.
No. The output follows the uploaded still, so the shape selector only appears in text-to-video mode.
Not on this platform yet. Text and one starting image are the available inputs today.
Other models on this platform that do a similar job, or that pair well with this one.
Wan 3.0 Prime is the second selectable product in the Wan 3.0 family: the same controls and length range as the standard variant, quoted at a higher credit rate.
Read the guideSeedance 2.5 creates a clip from a written description or from one starting image, with whole-second lengths from 4 to 30 seconds and three quality tiers.
Read the guideOne of the two selectable GPT Image 2.5 products on this platform: text and reference-based image generation with transparent-background support at the 1K tier.
Read the guide