Video
MiniMax H3
Video
Available on Ofrai — chat, images, and video in one place with shared credits.
Example prompts
Choose an example to load its prompt and recommended settings into the MiniMax H3 creator above.
How it works
Create with MiniMax H3 in three steps
- 01
Choose your inputs
Keep MiniMax H3 selected, then add a prompt and any supported reference assets.
- 02
Set the output
Choose the aspect ratio, resolution, and available Video options before generating.
- 03
Generate and refine
Review the result, adjust the prompt or settings, and generate another version from the same place.
The documentation page describing this endpoint is titled MiniMax Hailuo H3 video generation, and it documents three model ids: MiniMax-H3, MiniMax-H3-Max and MiniMax-H3-Max-Turbo. It does not define a fourth id named H3 Video, and this page does not invent one.
What follows is the Hailuo H3 video request contract exactly as the documentation states it, with the base MiniMax-H3 treated as the entry point because it is the one variant that still offers 2K, the one the documented default resolution belongs to, and the one that reaches down to four seconds where the Max tiers start at five.
If the label you arrived with names a Max tier, the parameters are still the same ones documented here, and the values you send for resolution, duration and reference media are what determine which tier renders your job.
What this endpoint is documented as, and what it is not
The documentation describes one video generation surface serving three model ids, not three surfaces and not four. The practical reading is that there is a single request shape and the tier is chosen by what you put in it. Text-to-video needs no media. One or two images attached with declared roles drive the frame modes.
Three to nine images drive the combined reference mode, and that one is unavailable on Max-Turbo, so a request that is valid against MiniMax-H3 or MiniMax-H3-Max will be refused by the Turbo id. Duration and resolution then place the request on the rest of the ladder: 4 to 15 seconds and 2K or 768p for the base tier, 5 to 15 seconds and 480P or 768P for the two Max tiers.
There is no mode parameter on this endpoint at all, which is worth stating plainly, because it means the shape of the request is the only control you have.
The base tier is the one that still reaches 2K
MiniMax-H3 accepts 2K and 768p for resolution, while MiniMax-H3-Max and MiniMax-H3-Max-Turbo accept 480P and 768P. On duration the base tier runs 4 to 15 seconds against 5 to 15 for the Max tiers. The endpoint's documented default resolution is 2K, so a request that names nothing lands at the top of the base tier's ladder rather than the bottom — that default is worth being deliberate about, because a draft that inherits it is a 2K draft.
The four-second floor on the base tier is the other half of the difference and the more useful half in practice: four seconds is the natural length of a stinger, a loop point, or a cut that has to fit an existing slot, and it is unavailable above. Note also that the documentation writes the mid rung with different casing on the two tiers, 768p for MiniMax-H3 and 768P for the Max tiers, so send the exact token rather than a normalised one.
One prompt, three generation modes
The family supports text-to-video, first-frame and last-frame image-to-video, and a combined mode in which image, video and audio references are used together — that last one is a single mode using several kinds of material, not three features to be enabled independently. Which mode you are in is implied by the inputs rather than declared: no images means text-to-video, one or two images with roles means the frame modes, three or more means the combined reference mode.
The aspect ratio field confirms this, because its behaviour changes with the mode. It defaults to 16:9 in text-to-video and cannot take adaptive there; it is ignored entirely in the frame modes because the source image governs the frame; and it defaults to adaptive in the combined reference mode while still accepting an explicit value. If a request is behaving oddly, checking which of those three states it is in resolves most cases.
The image count assigns the role
This is the mechanic that catches people on all three ids. The compatibility field image_urls takes a URL array and derives each image's purpose from its length: one image is the first frame, two are the first frame and then the last frame in order, and three to nine are reference images, which only MiniMax-H3 and MiniMax-H3-Max support. Nine is the ceiling.
New integrations are pointed at image_with_roles instead, which names each role explicitly, and the documentation warns specifically against sending both fields because the purpose of an image can end up ambiguous. The general rule to take away is that role should never be inferred on your behalf when the mode you are in makes it matter — a three-image request is a reference request, not a request with one image and two extras, and treating it as anything else produces output that looks like a mistake rather than an error.
Submitting returns a task, not a clip
The call is asynchronous and returns a generation.task object. That is the whole contract, and it has three consequences worth designing around. Reference media must already exist as reachable URLs at the moment of submission, because nothing in the request performs an upload for you. The response cannot be treated as a result, so polling has to be part of the integration.
And a failure in an asynchronous interface surfaces differently from a failure in a synchronous one: the request is accepted, then the job fails later, which means your error handling needs to cover the status transition as well as the response. None of this is exotic, but all of it is the difference between an integration that works on the first job and one that works on the tenth.
7000 characters, and the one field you cannot skip
prompt is required and capped at 7000 characters. The documentation's guidance on structure is the useful part: describe the subject, the action, the scene, the camera, the style and the sound requirements, in that order. In the combined reference mode the subject clause is where references belong, so the sentence order is load-bearing there — a reference named after the camera work tends to be read as scenery.
Seven thousand characters is far more than most shots need, and the risk of the large budget is not that you will run out of room but that you will use it. Two conflicting descriptions of the same subject produce a diluted result, not a richer one. Write the subject once, let the references hold identity, and keep the camera language in its own sentence so that a failed render can be diagnosed as either a reference problem or a camera problem.
FAQ
- Is MiniMax H3 Video a separate model id?
- Not according to the documentation. The documentation page is titled MiniMax Hailuo H3 video generation and defines three ids — MiniMax-H3, MiniMax-H3-Max and MiniMax-H3-Max-Turbo. There is no fourth id named H3 Video; a request is placed on a tier by its resolution, duration and reference media.
- What resolution and duration does the base MiniMax-H3 accept?
- Resolution is 2K or 768p, with 2K as the documented default for the endpoint, and duration is any whole number from 4 to 15 seconds. The two Max tiers accept 480P or 768P and 5 to 15 seconds, so a four-second request only fits the base tier.
- How is each attached image used?
- Through the compatibility field the count decides: one image becomes the first frame, two become first frame then last frame, and three to nine become reference images, which the base tier and MiniMax-H3-Max support but Turbo does not. Up to nine images. Declare roles explicitly with image_with_roles, and never send it together with image_urls.
- Can I ask for 9:16 when I am supplying a first frame?
- No — in the frame modes the output always adapts to the input image and the ratio is ignored. adaptive is unavailable in text-to-video, where the default is 16:9, and becomes the default only in the combined reference mode. In other words the aspect ratio is a constraint, a no-op, or a default depending on which inputs are present.
FAQ
- Can I use MiniMax H3 directly on this page?
- Yes. The creator above is the working Ofrai studio with MiniMax H3 selected by default.
- Where are my finished results saved?
- Signed-in generations are kept in your Ofrai library so you can download or revisit them.
- Which prompts work best with MiniMax H3?
- Describe the subject, setting, visual style, lighting, and camera or motion clearly. The examples above provide practical starting points.
- Can I edit the example prompts before generating?
- Yes. Load any example, then change its wording, aspect ratio, references, or other supported settings in the creator.
Related models
Compare other Video models that can handle similar creative work to MiniMax H3.
Related guides
Tools
Commercial use depends on the selected model and model terms.





