MiniMax H3
MiniMax H3 is a video model family by MiniMax. H3 connects generation, references, editing, motion transfer and replacing a hero in an existing clip in Vini. Choose the operation by what the source should control. Other services from the author are not implied to be available by this H3 page.
Variant, provider and resolution
Section titled “Variant, provider and resolution”H3 has two variants of the same model: fast (Max) (быстрое (Max)) and full (H3) (полное (H3)). The inputs are the same; the address, the price and the resolution list change. Since 23 October 2026 new nodes use fast (Max) for text-to-video and first-frame generation (a last frame is accepted). In From references, editing, motion transfer and Replace hero there is no variant choice yet: they run the full model.
New nodes send the request through fal. The Directly at MiniMax (напрямую у MiniMax) checkbox, off by default, sends it to MiniMax itself; the price and the resolution list change with it, and the card recalculates before the run. The checkbox is locked, with the reason shown, when only one provider carries the operation: Replace hero runs only through fal; From references, editing and motion transfer only directly at MiniMax. Since 25 October 2026 From references runs directly at MiniMax, so it offers 480P, 768P and 2K (no 4K).
| Provider | fast (Max) | full (H3) |
|---|---|---|
| fal | 480P, 768P, 1080P | 480P, 768P, 2K, 4K |
| directly at MiniMax | 480P, 768P | 480P, 768P, 2K |
New nodes use 480P with native sound. Nodes created before 23 October 2026 keep what they had: full H3 directly at MiniMax, the same request and the same cache signature, so their finished clips are not regenerated. The connected providers have no sound-off switch. The current studio duration remains 5–15 seconds; editing and motion transfer use their source-length rules below. H3 Max Turbo is not connected. The declared-limits table below is the registry projection for the MiniMax provider; the provider table above is what the card offers for each choice.
Replace hero (Заменить героя) is the one operation that runs through a different service route. It uses the H3 Max address at fal, which accepts only that operation; the MiniMax provider named above has no such operation. H3 Max here is therefore wiring for that role, not a quality level you choose, and it does not change the other roles. Since 23 October 2026 that route is the fal-ai/minimax-h3-max-recast address. Replace hero offers 768P (new nodes) and 1080P; the cost is shown before the run. It is the output rate for the source clip’s length plus a per-second charge for reading the source clip; on 768P the number of photos, the source frame size and the number of shots did not change the charged sum in our measurements (two people in the source; 1080P not measured).
In From references, pictures beyond the fifth add a per-picture charge; the card includes it before the run.
The MiniMax API reference and pricing were checked on 20 October 2026. Those public pages list 768P and 2K for H3; the studio’s 480P option is additionally supported by its earlier accepted runs and supplier billing records. The public documentation alone does not establish H3 480P support or a fresh quality result.
| Setting | Value |
|---|---|
| Model ID | minimax_h3 |
| Name | MiniMax H3 |
| Availability | available |
| Default for new nodes | no |
Common declared limits
Section titled “Common declared limits”| Setting | Value |
|---|---|
| Capabilities | text to video, image to video, dialogue, reference audio |
| Prompt required | yes |
| Maximum prompt length | 7000 |
| Duration | 5–15 seconds; fallback 5 seconds |
| Audio | model-generated audio; fallback model-generated audio |
| Aspect ratio | 9:16, 16:9, 4:3, 3:4, 1:1, 21:9; fallback 9:16 |
| Declared language support | prompt: English, Russian; speech: English, Russian; tier: silver |
| Controls | resolution: 480P, 768P; new node 480P; fallback 480P; quality: max, full; new node max; fallback full |
| Element input form | sheet |
| Content constraints | people: restriction level: ok |
| Shots | seconds per shot: 5; shot timing format: timecodes |
| Negative prompt strategy | instruction in prompt text |
Mode: from_first_frame
Section titled “Mode: from_first_frame”| Setting | Value |
|---|---|
| Aspect ratio | not declared |
| First frame | required |
| Last frame | accepted; optional |
Quality: full in from_first_frame
Section titled “Quality: full in from_first_frame”| Setting | Value |
|---|---|
| Controls | resolution: 480P, 768P, 2K; new node 480P; fallback 480P; quality: max, full; new node max; fallback full |
Mode: text_to_video
Section titled “Mode: text_to_video”Uses the common declared limits above.
Quality: full in text_to_video
Section titled “Quality: full in text_to_video”| Setting | Value |
|---|---|
| Controls | resolution: 480P, 768P, 2K; new node 480P; fallback 480P; quality: max, full; new node max; fallback full |
Mode: from_references
Section titled “Mode: from_references”| Setting | Value |
|---|---|
| Combined reference limits | images: 9; videos: 0; audio inputs: 3; combined maximum: 12; audio without visual input: no; audio reference seconds: minimum: 2; maximum: 15; combined maximum: 15; trim: head |
| Controls | resolution: 480P, 768P, 2K; new node 480P; fallback 480P; quality: not declared |
| Element input form | frames |
| References | maximum: 9 |
Mode: edit_video
Section titled “Mode: edit_video”| Setting | Value |
|---|---|
| Duration | same as source video; fallback same as source video |
| Aspect ratio | same as source video |
| Combined reference limits | images: 9; videos: 1; audio inputs: 0; combined maximum: 10; audio without visual input: no; video reference seconds: minimum: 2; maximum: 15; combined maximum: 15; video fps min: 23.976 |
| Controls | resolution: 480P, 768P, 2K; new node 480P; fallback 480P; quality: not declared |
| Shots | not available in this configuration |
| Source video | required |
| References | maximum: 9 |
Mode: motion_transfer
Section titled “Mode: motion_transfer”| Setting | Value |
|---|---|
| Prompt required | no |
| Duration | same as source video; fallback same as source video |
| Aspect ratio | same as source video |
| Combined reference limits | images: 9; videos: 1; audio inputs: 0; combined maximum: 10; audio without visual input: no; video reference seconds: minimum: 2; maximum: 15; combined maximum: 15; video fps min: 23.976 |
| Controls | resolution: 480P, 768P, 2K; new node 480P; fallback 480P; quality: not declared |
| Required media counts | images: 1 |
| Shots | not available in this configuration |
| Source video | required |
| References | maximum: 9 |
Mode: recast
Section titled “Mode: recast”| Setting | Value |
|---|---|
| Prompt required | no |
| Maximum prompt length | 2000 |
| Duration | same as source video; fallback same as source video |
| Audio | origin; fallback origin |
| Aspect ratio | same as source video |
| Combined reference limits | images: 4; videos: 1; audio inputs: 0; combined maximum: 5; audio without visual input: no; video reference seconds: minimum: 5; maximum: 30 |
| Controls | resolution: 768P, 1080P; new node 768P; fallback 768P; quality: not declared |
| Required media counts | images: 1 |
| Shots | not available in this configuration |
| Source video | required; explicit mode selection required; scene clip is not inserted automatically |
| References | maximum: 4; views per Element: 1 |
Unavailable modes: avatar_audio.
Limits apply together. initial is the studio choice for a new node; default is the fallback for an existing node without that value. Declared language or input support does not guarantee a successful result.
Prompting playbook
Section titled “Prompting playbook”- Give each visual or audio reference a specific role. Describe the relationship between the subject and the source motion.
- For motion transfer, describe the destination environment explicitly; it need not be the donor clip’s background.
- For a longer scene, keep actions sequential and preserve a clear starting pose.
The built-in MiniMax H3 authoring skill adapts a scene request for this family through Prompt for the model. It prepares instructions; it does not render video or certify the result.
Replace hero
Section titled “Replace hero”Take an existing clip and put different people into it: the background, camera, timing, gestures and the source sound stay, while the people are redrawn from your photos. Connect a source clip of 5–30 seconds (the provider rejects a single shot longer than 14.7 seconds, after the queue) and one to four Elements, one per person, in the order they stand in the frame from left to right. The prompt is optional (up to 2,000 characters). Photos should be at least 256 px.
The mode takes one picture per hero. If an Element carries several views, only one is sent and the studio shows a line saying so. Character names written as @name become plain words because this address has no reference tokens. Auto and a scene clip never select this mode, because it is the costlier route; choose it explicitly. The output is rendered at the source clip’s length, format and with its sound.
Studio example
Section titled “Studio example”Use a walking donor clip and an Element for a fictional traveler. Select H3 motion transfer. Ask Prompt for the model: “The traveler repeats the walking pace on a quiet beach. The coat and round glasses stay recognizable.” Check the model’s actual inputs and inherited format before running.
A run may consume credits. Review the prompt and inputs, then inspect the finished clip for action, identity, framing and sound. For automation, use the supported MCP workflow.
Known failure modes
Section titled “Known failure modes”A reference sheet can introduce layout or text into a scene; follow the mode’s current hero-input policy. Background signs may be garbled even when identity succeeds. Passing motion transfer does not prove exact pose-for-pose choreography.
Evidence and limits
Section titled “Evidence and limits”The 22 October 2026 exam of Replace hero used a six-second landscape clip with two people and two character sheets. Both people were replaced in the right order, and the background, gestures, cuts and source sound were kept. The vendor’s input schema was read on the same date. This is one case: it does not cover a single person, very long clips, close faces, fast action or other sources, and it does not establish the result for sheets versus clean frames beyond that run.
The older editing exam used individual frames. Editing now uses sheets and its replacement exam is outstanding; the older result does not certify the current editing input.
The 28–30 September 2026 exams retained the character in reference, edit and motion-transfer cases. In motion transfer, the requested beach replaced the donor setting, and output followed the donor’s landscape format. The edit evidence used clean character frames; it does not re-test a later sheet-based input arrangement.
The studio registry establishes available inputs and modes. A recorded render establishes only its observed outcome. The applicability guide explains the distinction. See also character identity, static camera and video style samples.
See the model–technique applicability matrix for connected operations, dated outcomes and configurations that remain untested.
Reviewed 14 October 2026 against the current studio contract and dated studio research. Official model-author documentation checked on that date; author capabilities may exceed Vini’s connected controls. Recheck this playbook when the model version, quality or mode changes.