Video
MiniMax H3 Max
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 Max creator above.
How it works
Create with MiniMax H3 Max in three steps
- 01
Choose your inputs
Keep MiniMax H3 Max 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 Hailuo H3 documentation covers three model ids, and MiniMax H3 Max is the one that keeps the whole reference set while sitting below the base tier's resolution ceiling. Max accepts the same generation modes as MiniMax-H3 — text-to-video, first-frame and last-frame image-to-video, and a combined reference run in which images, video and audio are submitted together — but its ladder stops at 768P with a 480P option underneath, and its duration window starts at five seconds instead of four.
Two consequences decide most real jobs. The default resolution documented for the endpoint is 2K, which belongs to MiniMax-H3 and is not one of Max's options, so a Max request has to name 480P or 768P itself. And in the combined reference mode Max is the tier where the default aspect ratio flips away from 16:9 and becomes adaptive, because the frame is derived from the references rather than chosen up front. Max is a reference tool with a shorter resolution range, not a smaller reference tool.
Where Max sits inside a three-id family
The reference lists MiniMax-H3, MiniMax-H3-Max and MiniMax-H3-Max-Turbo, and the capability split is not evenly drawn. MiniMax-H3 and MiniMax-H3-Max both support text-to-video, first-frame and last-frame image-to-video, and combined generation from image, video and audio references. Max-Turbo supports only the first two: reference media are not available to it at all, including the three-to-nine-image reference set.
So the family divides on whether reference media exist, and Max is one of the two ids where they do. What separates Max from the base tier is its ladder rather than its inputs: 480P and 768P against 2K and 768p, and 5 to 15 seconds against 4 to 15. That distinction is not cosmetic. If the reference set is the point of the job, Max delivers the full set at a resolution that has to be graded before delivery rather than one that does not.
If the 2K master is the point, the base tier is the correct request, and Max is the wrong one at any level of prompt quality.
Max's resolution and duration envelope
Max's resolution field takes 480P or 768P, and nothing else. The documented default for the endpoint is 2K, which is one of MiniMax-H3's two options and not one of Max's, so treat resolution on a Max request as a value you set rather than one you can leave out. Duration on Max is any whole number from 5 to 15, and no default duration is documented, which points the same way: name it. The practical differences are worth internalising.
The five-second floor means four-second inserts — stingers, a loop-able cut, a bumper that has to land inside a shorter slot — are simply not available on Max, while the base tier reaches down to four. At the other end both tiers reach 15, so long work is not a Max constraint.
The integer granularity is the genuinely useful part: every duration in 5 through 15 is valid rather than a short list of buckets, which lets a shot be matched to a beat or fitted to an existing edit slot instead of padded out from a five-second render.
image_with_roles: the input list that assigns its own roles
The compatibility field image_urls takes a plain array of URLs and infers meaning from how many entries there are. One image becomes the first frame. Two become the first frame and then the last frame, in that order. Three to nine become reference images, and that third case exists only on MiniMax-H3 and MiniMax-H3-Max — Max-Turbo cannot take it. Nine reference images is the hard ceiling for the whole family.
New integrations are steered away from the count-inference field and towards image_with_roles, which states each image's purpose explicitly instead of letting an array length decide it, and the documentation is blunt about the failure mode: sending image_urls and image_with_roles together can leave the purpose of an image ambiguous. On Max this matters most, because Max is where a three-or-more-image reference job is actually likely to happen.
If the look is being assembled from a set of stills, letting a padded list infer roles is a bad bargain.
The same aspect ratio answers three different questions
The field behaves differently depending on which other inputs are present, and this is the single most common source of confusion in this family. In text-to-video the default is 16:9 and adaptive is not available at all. In first-frame and last-frame mode the output always adapts to the input image, and any ratio passed in is ignored — a landscape first frame with 9:16 requested does not letterbox or crop, the value is discarded.
In the combined reference mode, available on MiniMax-H3 and MiniMax-H3-Max only, the default is adaptive and a specific ratio can still be named. Seven values are listed overall: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and adaptive. So the field is a constraint, a no-op, or a default depending on the mode, and what decides the mode is which image inputs you attached and how you declared their roles.
What 7000 characters of prompt actually buys
prompt is the one required field in this API, capped at 7000 characters. The cap is the less interesting fact. The requirement is the interesting one, because it means there is no generate-without-describing-it entry point: every job on Max explains itself in words before anything is rendered. The documentation's own structure suggestion is worth following literally — subject, action, scene, camera, style and sound requirements, named in that order.
In the combined reference mode the subject clause is exactly where the reference indices belong, so the order of the sentences decides whether the model reads your references as the subject or as scenery. Seven thousand characters is a very large budget in either language, and it changes the failure mode rather than removing it: prompts that short fail by being vague, and prompts this long fail by containing two competing descriptions of the same thing. One subject, one camera, and let the references carry identity.
Submitting is not rendering: the async task contract
The endpoint is asynchronous. A submission returns a generation.task object rather than a finished clip, and that shape has concrete consequences for how the integration is built. Reference media must already be reachable when the call goes out, because the request carries URLs rather than performing an upload — anything that has to be uploaded has to be uploaded first and its URL collected.
A call that returns a task identifier cannot be evaluated inline, so status polling belongs in the pipeline from the first version rather than being bolted on when a batch job hangs. The nine-image ceiling sits alongside this as a hard boundary rather than a soft preference: if the brief implies a tenth reference, something has to be dropped or substituted before the request is made, because the failure will not be a silently degraded image.
FAQ
- How is MiniMax H3 Max different from the base MiniMax-H3?
- The reference lists both as supporting text-to-video, first-frame and last-frame image-to-video, and combined image, video and audio reference generation, so their inputs match. They differ on the ladder: Max is 480P and 768P with 5 to 15 seconds, while the base tier is 2K and 768p with 4 to 15 seconds. Max-Turbo is the third id and drops reference media entirely.
- Do I need to set resolution and duration on a Max request?
- Set both. The documented default resolution is 2K, which belongs to MiniMax-H3 and is not a Max option, so Max needs 480P or 768P named explicitly. No default duration is documented either; Max accepts whole seconds from 5 to 15, so a value outside that range is a different model's window rather than a Max one.
- How many reference images does Max take, and what are they used for?
- Up to nine. Through the compatibility field, one image is used as the first frame, two become first frame then last frame, and three to nine become reference images — a case Max supports but Max-Turbo does not. New integrations should declare roles explicitly with image_with_roles, and the two fields must never be sent together.
- Can I request a 9:16 aspect ratio alongside first and last frame images?
- Not usefully. In first-frame and last-frame mode the output always adapts to the input image and a ratio passed in 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 supported by MiniMax-H3 and MiniMax-H3-Max.
FAQ
- Can I use MiniMax H3 Max directly on this page?
- Yes. The creator above is the working Ofrai studio with MiniMax H3 Max 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 Max?
- 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 Max.
Related guides
Tools
Commercial use depends on the selected model and model terms.





