MiniMax video model · open weights · 768P and 2K

Hailuo H3 Video Generator

Hailuo H3 is the search name for MiniMax H3, the generation whose weights MiniMax published. This page runs one image-to-video route on it, quotes the official per-second rates, and separates the model document from the controls here.

Try Hailuo H3
01

Create with Hailuo H3

Upload one frame, describe the motion, set duration and resolution, then read the point estimate before you submit. The official model accepts far more than this route exposes, and the difference is listed below rather than hidden.

LIVE
5s
2s15s

Results

Creation history

Generated content will appear here
02 / Overview

Hailuo H3 specifications

Facts reviewed:

Developer
MiniMax
Official model name
MiniMax H3
Official API model ID
MiniMax-H3
Alias in the official OpenAPI document
Hailuo-03
Release and open weights
31 July 2026, weights published 3 August 2026
Official output resolution
768P and 2K
Official clip length
4 to 15 seconds, whole seconds only
Official frame rate
24 fps
Official audio
32 kHz stereo, generated in the same pass as the picture
Stable dialogue languages
11
Weights licence
MiniMax H3 Community License, with excluded territories
Platform route
hailuo-3

What the model documents, and what this page actually exposes

The official tables describe a general-purpose omni-modal system. This page runs a single image-to-video route on it, and the two boundaries sit further apart here than on most of the other video pages, so they are listed separately rather than merged into one list.

OFFICIAL

Documented for MiniMax H3

  • Text-to-video from a prompt alone, with camera terms such as pan, zoom, and static written after the key description
  • Image-to-video from a first frame, a last frame, or both in the same request
  • Omni-reference generation combining up to 9 images, 3 video clips, and 3 audio clips
  • Native 32 kHz stereo audio produced in the same pass as the picture, with 11 stable dialogue languages
  • Output at 768P or 2K, 24 fps, from 4 to 15 whole seconds, snapping to a 17k plus 5 frame grid
  • Six fixed aspect ratios plus adaptive, which is forced for image-to-video because the input frame decides the ratio
  • Instruction-led editing and motion transfer between clips
  • Two separate tasks around generation: H3-Context-IR for prompt enrichment, and 768P-to-2K regeneration
  • Open weights for the two base checkpoints, released under a licence that excludes four territories
LIVE PLATFORM

Exposed by this workspace today

  • Image-to-video from one uploaded frame; no text-to-video and no first-and-last-frame route
  • A 2 to 15 second selector with 5 seconds as the default
  • 720p and 1080p in the resolution selector, neither of which is the documented 768P or 2K
  • Five fixed aspect ratios with 16:9 as the default
  • Reference-to-video appears as a selectable mode with no asset upload control behind it
  • A point estimate quoted before submission, and no generated-audio switch on this route

The gaps run in both directions rather than one way, so they are stated by field. Resolution moves inward and outward at once: the selector offers 1080p, which no official page lists for this model, while the documented 2K output is not available here. Duration moves the same way, with a 2-second floor below the official 4 seconds. Audio is a straight subtraction, because the model produces 32 kHz stereo in the same pass and this route exposes no audio control at all. Reference generation is the widest gap: the model accepts twelve reference files across three modalities, and this route shows the mode without offering a single upload slot.

Model brief

Where Hailuo H3 sits in the MiniMax video line

Hailuo H3 is the name people search for; MiniMax H3 is the name MiniMax uses; MiniMax-H3 is the only string that works in code. The official model list registers it as MiniMax H3 next to MiniMax H3 Max, the OpenAPI document for the video endpoint titles itself Hailuo-03, and the launch post from 31 July 2026 introduces it as a general-purpose multimodal generation model. The Hailuo prefix has an older history than the H3 suffix suggests, because the official model list uses Hailuo only for the previous generation, where MiniMax Hailuo 2.3, MiniMax Hailuo 2.3Fast, and MiniMax Hailuo 02 still sit. Treating Hailuo H3, MiniMax H3, and Hailuo-03 as three different products is the most common mistake in third-party coverage of this release.

What MiniMax actually shipped is a single system built around one context rather than a pipeline of specialists. The launch post states that H3 understands unified context across text, images, video, and audio, and generates video with native stereo sound, up to 15 seconds at 2K resolution. The system document splits that into three parts: H3-Context-IR, which interprets and restructures the multimodal instruction; H3-Base, which generates video and audio at a 768-pixel short edge; and H3-Regenerate-2K, which feeds the 768p result back together with the original context. The reason for the third part is stated plainly in the announcement, and it is not the usual one. Rather than attaching a dedicated super-resolution module, MiniMax has the base model regenerate its own low-resolution output in context, so that details the original context contains, such as small on-screen text, can be recovered instead of guessed.

The open release is the part of this release that has no counterpart on the other video pages here. MiniMax published weights for the two base checkpoints, H3-Base FL2VA for text-to-video and first-and-last-frame work, and H3-Base Ref2VA for reference-driven generation, and the official self-hosting guide documents ComfyUI templates from version 0.30.0 onward, a preview SGLang path from version 0.5.17, and a community vLLM-Omni recipe, alongside a LoRA fine-tuning route. Two boundaries come with that download and both are official. The open release does not contain H3-Context-IR or H3-Regenerate-2K, so a self-hosted service cannot reproduce the platform workflow end to end. And the licence names excluded territories, the United States, the European Union, the United Kingdom, and South Korea, which need a separate licence as of the revision the guide reviewed on 26 August 2026.

This page runs the platform route hailuo-3, and the honest summary of that route is short: it does one thing. Image-to-video with a single uploaded frame is the only path with controls behind it. The official model documents text-to-video, first-and-last-frame control, and reference generation across three modalities, so the capability list below is a description of the model rather than an inventory of buttons. Resolution is the same kind of statement. The selector here offers 720p and 1080p, and neither is a documented output for this model, whose official options are 768P and 2K. Duration is exposed from 2 to 15 seconds against an official 4 to 15, and there is no audio switch even though the model generates 32 kHz stereo in the same pass as the picture. Read the official sections as the ceiling, and the workspace sections as the floor you will actually work on.

03 / Capabilities

What the H3 generation changed

These are the points the official documents argue for directly, not the ones third-party pages repeat.

01

Audio arrives with the picture, not after it

Every H3 output carries native 32 kHz stereo generated in the same pass as the video, so dialogue, room tone, and effects are produced together rather than assembled afterwards. The announcement describes the audio encoder as processing the left and right channels independently and recombining them, and compressing 32 kHz audio into latent tokens at 40 Hz. On the price page, reference audio is the one input class that costs nothing.

02

Three task families collapsed into one model

The launch post walks through what it calls siloed tasks and explains the design goal as unifying and generalising across them. Voice, sound effects, and music are jointly modelled rather than separated; text-to-image, image-to-image, image-to-video, audio-to-audio, and audio-video-to-audio-video reference work are handled by one system. MiniMax also states that it set aside the Hailuo 02 architecture for H3, on the argument that architectural tricks should give way to how the model is defined.

03

2K comes from regeneration, not upscaling

The 2K path reuses the base model rather than attaching a super-resolution head. H3-Regenerate-2K takes the 768p result together with the original multimodal context and produces the higher-resolution output from both, which MiniMax presents as the reason fine detail survives. On the platform API you can request 2K directly; in the open-weight pipeline the second stage is required, and that second stage is not part of the published weights.

04

The weights are downloadable, with a stated boundary

Two checkpoints are published on Hugging Face and ModelScope, both CFG-distilled, both carrying their own processor, tokenizer, text encoder, visual VAE, and a standalone audio VAE. Official paths cover ComfyUI templates, an SGLang service with a reference topology, and a community vLLM-Omni recipe, plus LoRA fine-tuning. The licence qualifies that access by territory, which is a fact worth reading before planning a deployment rather than after.

05

Reference capacity measured in files, not in prompts

Reference generation accepts up to 9 images, 3 video clips, and 3 audio clips, capped at 12 files in total. Individual clips run from 2 to 15 seconds with a 15-second ceiling across the whole set, and every file carries its own format and size limit. Audio is the exception that proves the rule, because reference audio must be accompanied by an image or a video and cannot be the only input.

06

Six aspect ratios and one forced adaptive setting

The model supports 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, with adaptive available as a default rather than a seventh ratio. The rules change by mode: for text-to-video the ratio is required and cannot be adaptive, for image-to-video the input frame decides and the value is always adaptive, and for reference generation it is optional.

07

The H3 channel is provisioned differently from the old one

MiniMax publishes rate limits per API, and the contrast is inside one table. The V2 video endpoint serving MiniMax-H3 is listed at 300 requests per minute with a maximum of 30 in-flight tasks. The older video generation endpoint covering the Hailuo series is listed at 20 requests per minute with no in-flight figure published. The H3 channel is documented at fifteen times the request rate of the channel it replaces.

08

You can read the technical argument instead of a summary of it

MiniMax published the reasoning behind the architecture, including the captioning pipeline that expands multimodal context into roughly 100K tokens of inference and distils it to about 4K, a rebuilt tokenizer credited with a four-fold gain in effective sequence length, and a split between understanding and generation workloads that the company says lifted training throughput by close to 30 percent. The same post states that the full technical report is still to come and lists what H3 still needs to improve.

04 / Scene ideas

Jobs that fit the H3 generation

These are the points the official documents argue for directly, not the ones third-party pages repeat.

Spots where the soundtrack is part of the deliverable

Because audio is generated with the picture, a first pass already contains dialogue, ambience, and effects in stereo. That suits advertising and brand work where the sound design is reviewed alongside the shot rather than added in a later stage.

Dialogue in more than one language

Eleven languages are documented as stably supported for dialogue, including Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. MiniMax notes that other languages work to varying degrees, which is a reason to test rather than assume.

Motion transfer from an existing clip

The launch post names V2V motion transfer among the areas early testing favoured, and the reference route accepts video clips as inputs. That makes it usable where the movement is already decided and only the subject needs to change.

Frames where on-screen text has to survive

MiniMax lists accurate text and brand rendering among the strengths of this generation, and explains the 2K path partly by the recovery of small text. Any job that ends with legible type in frame is worth judging at full resolution before it is accepted.

Teams that need to run the model themselves

This is the only video model in the catalogue here with published weights and official self-hosting documentation. Where data cannot leave a controlled environment, or where the volume makes hosted per-second billing impractical, H3 offers a path the rest of the catalogue does not.

Reference-led character and product work

The reference route is designed to hold identity across a sequence by pinning appearance to supplied images, motion to a supplied clip, and voice to supplied audio. Each reference carries a declared role, which is what keeps four inputs from collapsing into a blend.

Transitions that must land on an exact frame

First-and-last-frame input turns a static pair into a controlled move, which is the natural fit for a product reveal or a shot that has to end on a specific composition. The official example for this mode is a two-image request with no reference arrays.

Iteration cycles that start from a description

The H3-Context-IR task exists to turn a loose multimodal brief into a structured prompt before generation, and MiniMax recommends putting it in the pipeline rather than treating it as optional. For work that runs many prompts against one setup, that pre-processing step is where quality is won or lost.

05 / Prompt lab

Prompt starters for the route this page exposes

Each of these is written for the single image-to-video path available here, so the prompt assumes one uploaded frame and describes movement rather than appearance. The official guide recommends putting camera terms such as pan, zoom, or static directly after the key description, and one camera movement per shot.

SCENE 01image to video

Portrait with a single motivated move

Hold identity and let one small movement carry the shot.

View prompt
Keep the uploaded person unchanged in face, hair, wardrobe, and the direction of the existing light. Add one slow turn of the head toward camera and a gentle breath, nothing else. Camera: a five percent push in, static afterwards. Keep the background layout and the shadow on the wall exactly as supplied. Quiet room tone only, no music, no added objects, no captions.
SCENE 02image to video

Product still, brought to life

Use one supplied studio frame and move only what has to move.

View prompt
Preserve the uploaded product shape, label position, and reflection on the surface. Camera: static, no movement at all. A single hand enters from the right, lifts the lid, and withdraws. Keep the label legible and unchanged, keep the highlight along the left edge, and add nothing to the background. Soft workshop ambience, one metallic click as the lid opens, no voice.
SCENE 03image to video

Environment reveal on a locked-off frame

Check how the model handles atmosphere and light when the camera does not move.

View prompt
Keep the uploaded street frame identical in perspective, colour, and signage. Camera: static. Fog thickens from the far end toward the camera, the nearest lamp flickers once and steadies, and one cyclist passes through the middle distance from left to right. Hold the geometry of the buildings unchanged. Distant traffic and a light rain bed, no dialogue, no text overlays.
SCENE 04image to video

Fabric and material study

A motion test that makes drift and weight easy to judge.

View prompt
Keep the uploaded fabric frame unchanged in colour, weave, and the fold pattern already present. Camera: slow lateral track to the right, no cut. A steady draught lifts the near edge of the fabric and lets it fall back, then lifts it a second time more gently. Keep the seams and the stitching identical throughout. Light indoor air, a faint cloth rustle, no music, no people.
SCENE 05image to video

Architectural interior with a walk-through

Use one supplied frame of a space and let the camera do the describing.

View prompt
Preserve the uploaded interior exactly in proportion, materials, and the position of every source of light. Camera: a steady forward push along the centre line, no pan, no cut. Dust drifts through the shaft of light from the left window, and a curtain settles at the far end. Keep the floor line and the ceiling height unchanged for the whole clip. Empty room reverb, no voices, no added furniture.
SCENE 06image to video

Character beat with a held expression

Test whether a supplied face survives a performance rather than a pose.

View prompt
Keep the uploaded character identical in facial structure, skin texture, hairline, and costume. Camera: static, medium close shot. The character listens, blinks once, looks down and then back up, and the expression shifts from guarded to relieved without any large movement. Keep both ears visible and both shoulders level. Close indoor ambience, one soft breath, no music, no subtitles.
COMPARE

Hailuo H3 and Hailuo H3 Max on one page

Both models are listed on the same official model overview, described in the same video generation guide, and priced in the same pay-as-you-go table. Every row below is read from those documents rather than assembled from third-party comparison articles.

Hailuo H3 and Hailuo H3 Max on one page
ComparisonHailuo H3 (MiniMax H3)MiniMax H3 Max
Official model nameMiniMax H3MiniMax H3 Max
API model IDMiniMax-H3MiniMax-H3-Max
Built byMiniMaxMiniMax and fal.ai, post-trained by fal.ai on MiniMax H3
Stated design goalGeneral-purpose omni-modal generationHigh-speed generation
Text-to-videoSupportedSupported
First-frame and last-frame image-to-videoSupportedSupported
Reference generationUp to 9 images, 3 videos, 3 audio clipsNot supported; the guide lists it as coming soon
Output resolution768P and 2K480P and 768P, with no 2K
Clip duration4 to 15 whole seconds5 to 15 whole seconds; 4 seconds is not accepted
Frame rate24 fps24 fps
Prompt limit per text item7,000 characters7,000 characters
Price per second, entry tier0.08 US dollars at 768P0.05 US dollars at 480P
Price per second, higher tier0.13 US dollars at 2K0.08 US dollars at 768P
Open weightsPublished for two base checkpointsNot published
LIVE PRICING

How MiniMax H3 is billed

MiniMax charges by the second of output, not by the request. On the pay-as-you-go table MiniMax-H3 is listed at 0.08 US dollars per second at 768P and 0.13 US dollars per second at 2K, while MiniMax-H3-Max is listed at 0.05 per second at 480P and 0.08 per second at 768P. Inputs are priced separately and asymmetrically: reference audio is free, the first five reference images are free with each additional image at 0.04 US dollars, and reference video is billed on its input duration at the same per-second rate as the output resolution. Regenerating a 768P clip to 2K costs 0.05 US dollars per second of regenerated output, and the price page notes that the original inputs are billed again for that pass. Prompt enrichment through H3-Context-IR is metered in tokens at 0.90 US dollars per million input tokens and 3.60 per million output tokens. MiniMax also states that H3 and H3 Max are reached through the pay-as-you-go API. This site does not bill in dollars; the workspace quotes platform points instead, and the live estimate before submission is the price that applies here.

Open the workspace to see the current estimate
01

Platform points

The workspace quotes points before you submit. Points are set by this site and are not the official per-second rate.

02

Resolution and duration on this route

MiniMax documents 768P and 2K from 4 to 15 whole seconds. This workspace exposes 720p and 1080p from 2 to 15 seconds, so the resolution selector uses a different naming scheme, 2K is unavailable here, 1080p has no official counterpart for this model, and the duration floor sits two seconds below the documented minimum.

06 / Workflow

Running an H3 generation end to end

  1. 01

    Decide the mode from what you can attach

    On the official endpoint the mode follows the media in the request rather than a separate switch: a prompt alone is text-to-video, a frame with a role is image-to-video, and reference assets select reference generation. On this route only the single-frame path has inputs behind it.

  2. 02

    Respect the one hard combination rule

    Image-to-video and reference generation cannot be mixed. If any reference image, video, or audio role appears in a request, first-frame and last-frame roles must not, and the reverse holds too. Planning around that rule is cheaper than discovering it at submission.

  3. 03

    Write the camera instruction after the description

    The official guide directs you to place camera terms such as pan, zoom, or static immediately after the key description, and to keep one camera movement per shot. That is what makes an image-to-video result read as deliberate rather than drifting.

  4. 04

    Set duration as a whole number inside the documented window

    Duration is an integer from 4 to 15 seconds on MiniMax-H3 and 5 to 15 on H3 Max, and official documentation states that the value snaps to a 17k plus 5 frame grid at 24 fps. The selector on this page starts at 2 seconds, which the model does not accept, so keep the setting at or above 4.

  5. 05

    Submit, poll, and collect the URL

    Generation is asynchronous. The create call returns a task ID, the query endpoint returns status and, on success, a direct download URL for the video. MiniMax recommends polling every ten seconds, and the terminal failure states are failed and cancelled.

Review before publishing

Boundaries, licence terms, and the gap between weights and workflow

A larger capability list does not remove the usual failure modes. Identity drift across a long clip, unstable geometry, an unintended cut, on-screen text that almost reads, audio that separates from the picture, and hands or props losing their shape can all still appear. Generate at the lowest cost setting your judgement allows, watch the whole clip including the soundtrack before accepting a take, and keep the prompts that worked, because this model responds to explicit camera direction and to the same subject described with the same words each time.

The two service boundaries on this page are not interchangeable and they move in opposite directions, which is why they are stated field by field. Resolution narrows and widens at once: the selector here offers 1080p, which no official page lists for this model, while 2K is documented but not available on this route. Duration narrows in one direction only, with a 2-second floor here against an official minimum of 4. Audio is a subtraction rather than a difference, because the model produces 32 kHz stereo in the same pass and this route exposes no audio control. Reference generation is the widest gap of all: the model takes up to 12 reference files across images, video, and audio, and this route presents the mode without a single upload slot behind it.

The licence is part of the product, not small print attached to it. The weights are released under the MiniMax H3 Community License, and the official self-hosting guide names the United States, the European Union, the United Kingdom, and South Korea as excluded territories requiring a separate licence, as of the revision reviewed on 26 August 2026. Anyone planning to download, deploy, or offer a hosted service should read the current agreement rather than this page, because a licence revision is not something a review date can freeze. Separately, the usual synthetic-media obligations apply: only upload material you are authorised to use, do not impersonate real people, and disclose generated footage where the platform or the jurisdiction requires it.

Downloading the weights does not reproduce the platform. MiniMax states that the open release contains the two H3-Base checkpoints and does not contain H3-Context-IR or H3-Regenerate-2K, so a self-hosted service cannot run the full prompt-enrichment and 2K workflow that the hosted API performs. The hardware picture is documented but deliberately unquantified on the point that matters most: the reference topology is a single node with eight B200 GPUs, the upstream cookbook reports roughly 108 GB of weights, and MiniMax states explicitly that this is a reference configuration rather than a minimum, and that peak memory and latency for that exact recipe are not published. The official guide also warns against mixing benchmark numbers between the quantised ComfyUI conversions and the official mixed-precision checkpoint, because they are different artefacts with different quality and memory behaviour.

Sources

Sources and how these facts were checked

Model identity, specifications, limits, licence terms, and pricing come from MiniMax documentation and MiniMax announcements, read at the source rather than through third-party summaries. Platform controls, points, and availability come from this site. Where the two differ, the difference is written out instead of smoothed over, and every price and promotional statement carries the date it was read.

07 / FAQ

Hailuo H3 questions, answered

Is Hailuo H3 the same model as MiniMax H3?
Yes. MiniMax H3 is the official model name in the company model list, MiniMax-H3 is the model identifier the video endpoint accepts, and Hailuo-03 is the alias used inside the official OpenAPI document for the same endpoint. Hailuo H3 combines the consumer brand name with the version number, which is why it became the common search term. Note that the official model list uses the Hailuo prefix only for the previous generation, where MiniMax Hailuo 2.3, MiniMax Hailuo 2.3Fast, and MiniMax Hailuo 02 are still registered.
Is Hailuo H3 open source?
The weights are published. MiniMax announced the open-source release on 3 August 2026, after launching the model on 31 July 2026, and published two checkpoints: H3-Base FL2VA for text-to-video and first-and-last-frame work, and H3-Base Ref2VA for reference-driven generation. Both are CFG-distilled models with their own processor, tokenizer, text encoder, visual VAE, and audio VAE. What is not included is as important: the open release omits H3-Context-IR and H3-Regenerate-2K.
Can the weights be used in the United States or the European Union?
The official self-hosting guide names the United States, the European Union, the United Kingdom, and South Korea as excluded territories that require a separate licence, as of the agreement revision the guide reviewed on 26 August 2026. MiniMax also frames the open release as subject to applicable laws and regulations, and advises reviewing the current agreement before downloading, deploying, or offering a hosted service. Check the licence itself rather than relying on any dated summary, including this one.
Does Hailuo H3 generate audio?
Yes, and it generates it in the same pass as the picture rather than dubbing a finished edit. Output audio is 32 kHz stereo, and the official architecture description explains that the audio encoder processes the left and right channels independently before recombining them. Dialogue is documented as stably supported in eleven languages, which are Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. MiniMax notes other languages work to varying degrees but publishes no number for them.
What resolution, duration, and frame rate does it output?
MiniMax-H3 outputs 768P or 2K at 24 fps, in whole-second clips from 4 to 15 seconds. The default short edge is 768 pixels with dimensions aligned to a multiple of 32, and the native 16:9 canvas is 1344 by 768. MiniMax also documents the pixel-area ceiling of 768 by 1344 for that tier. H3 Max is the narrower model here, at 480P or 768P and 5 to 15 seconds, with 2K explicitly not supported.
How much does MiniMax H3 cost per second?
The pay-as-you-go table lists MiniMax-H3 at 0.08 US dollars per second at 768P and 0.13 US dollars per second at 2K. MiniMax-H3-Max is listed at 0.05 per second at 480P and 0.08 per second at 768P. Inputs are separate: reference audio is free, the first five reference images are free with each additional image at 0.04 US dollars, and reference video is billed on input duration at the output resolution rate.
Is the price here the same as the official per-second rate?
No, and the two should not be compared as if they were. MiniMax bills in US dollars per second on its platform. This site bills in platform points, quoted in the workspace before you submit, so the number you see here is this site price rather than the official rate. Note also that MiniMax runs a separate membership-based pricing system on its Chinese consumer surface, which uses different figures and its own plan tiers; this page does not quote that system, because mixing it with the API table is how incorrect price comparisons get published.
How does the model produce a 2K output?
Through context-based regeneration rather than a super-resolution module. H3-Regenerate-2K feeds the 768p result back together with the original multimodal context, so the higher-resolution pass can recover small text and fine detail that upscaling would only approximate. On the platform API you can also request 2K directly when creating a task, with resolution set to 2K and no second step. The regeneration route is priced separately at 0.05 US dollars per second of regenerated output, and the original inputs are billed again for that pass.
What is H3-Context-IR and do I need it?
It is a separate task that interprets multimodal context across text, images, audio, and video, reasons about how those inputs relate, and returns a structured prompt with more semantic detail. It does not generate video. MiniMax describes it as critical to the quality of the final output and strongly recommends incorporating it into a generation pipeline, either by calling the endpoint or by building equivalent context processing. It is metered in tokens, at 0.90 US dollars per million input tokens and 3.60 per million output tokens, and it is not part of the open-weight release.
What is the difference between MiniMax H3 and MiniMax H3 Max?
H3 Max is the faster, narrower model. MiniMax describes it as jointly released with fal.ai and post-trained by fal.ai on MiniMax H3, optimised for high-speed generation, with a stated design goal of speed rather than general-purpose omni-modal work. It drops reference generation entirely, caps resolution at 768P with 2K explicitly unsupported, and raises the duration floor to 5 seconds. It is cheaper per second at matched lower resolutions, at 0.05 for 480P and 0.08 for 768P, and its weights are not published.
Can I combine a first-frame image with reference images in one request?
No. The official endpoint states the two are mutually exclusive: if any reference image, reference video, or reference audio role appears in the content array, first-frame and last-frame roles must not appear, and the reverse applies as well. That single rule decides several production questions up front, because it means a shot cannot be anchored to an exact opening frame and simultaneously built from character and motion references.
Which generation routes does this page actually run?
One: image-to-video from a single uploaded frame. Text-to-video, first-and-last-frame control, and reference generation are documented model capabilities that this route does not expose, and the reference mode appears as a selectable option with no upload control behind it. Resolution here is 720p or 1080p rather than the documented 768P and 2K, duration starts at 2 seconds rather than 4, and there is no audio switch even though the model generates native 32 kHz stereo.

Other video models on this platform

01

Seedance 2.5

Compare a closed flagship route with longer clips and 1080p at 10-bit.

02

Kling 3.0 Turbo

Compare another text-and-image route with a different provider boundary.

03

Grok Imagine Video 1.5

Compare a route where 1080p is documented for some workflows and not others.

04

Image to Video

Open the general image-led video workspace with model selection.

Render your next scene

Start a Hailuo H3 generation

Return to the live workspace, upload one frame, keep the duration at or above four seconds, and review the current point estimate before you generate.

Open Hailuo H3 workspace