Official model guide and online generator
Generate 2K video with native stereo sound from text, frames, and up to fifteen reference images, clips, and audio tracks at once. Use the MiniMax H3 AI video generator directly below to create from a prompt, image, frame pair, or supported references.
The complete Musci3 video workspace is embedded here and starts with MiniMax H3 selected. You can still compare variants or switch models without losing the rest of the workflow.
Generation requires an account and uses credits based on the selected model, variant, duration, resolution, and other settings. Failed tasks are refunded automatically.
MiniMax H3, released as Hailuo 3.0 in the Hailuo app, is a general-purpose omni-modal generation model rather than a text-to-video model with reference features bolted on. Text, images, video, and audio enter as one context, and the model returns picture and 32 kHz stereo sound generated together. Earlier video stacks split this work across separate expert models for text-to-video, image-to-video, first-and-last-frame, subject reference, motion reference, and editing; H3 folds them into a single model that takes the relationship between your references and your shot as plain language.
Native output at 24 fps
Clip length, in whole seconds
Reference images, clips, and audio per generation
Native stereo sound on every take
Omni-reference
This is MiniMax's own demonstration, and it is the clearest statement of what separates H3 from a text-to-video model with a reference slot. Three files of three different kinds went in, along with one sentence describing how they relate to each other. No mode switch, no separate motion-transfer model, no post-production pass to add the voice.
The references
“Reference the Hitchcock camera movement from Video 1, have the character in Image 2 sing, with the vocals matching Audio 3.”
Capabilities
H3 generates at 2K rather than enlarging a smaller result, so skin texture, fabric weave, reflections, and small on-screen type stay readable when the clip fills a display instead of a feed thumbnail. That matters most on the shots you were previously forced to reshoot or retouch: close-ups, product surfaces, and anything with legible detail near the edge of frame.
Sound is not a second job here. H3 predicts audio and video latents together and returns 32 kHz stereo, with stable dialogue across eleven languages, so lip movement and vocals are generated against each other rather than aligned afterwards. Every generation on this page comes back with its track already attached — there is no audio toggle to forget.
The system runs in three stages. H3-Context-IR reads your free-form multimodal input and rewrites it as a structured context. H3-Base generates picture and sound together at 768p. H3-Regenerate-2K then feeds that result back through the model along with the original context, so the higher resolution is generated from the scene's own information instead of interpolated on top of it.
Both models are on Musci3. Hailuo 2.3 remains the cheaper choice for expressive single-shot motion; H3 is the one to reach for when a shot needs sound, references, or delivery resolution.
| Capability | Hailuo 2.3 | MiniMax H3 |
|---|---|---|
| Generation modes | Text and image to video | Text, image, first-and-last-frame, and omni-reference |
| Resolution | 768p and 1080p | 768p and native 2K |
| Duration | 6 or 10 seconds | Any whole number from 4 to 15 seconds |
| Audio | Silent; sound designed in post | Native 32 kHz stereo, generated with the picture |
| Reference inputs | One starting image | Up to 9 images, 3 video clips, and 3 audio clips |
| Aspect ratios | 16:9 landscape | 21:9 through 9:16 |
Each clip below is MiniMax's own example for that category, generated with H3 and re-encoded here. Open the generator above to work in the same modes.
Build title sequences that cut between shots and carry their own score and sound design, at delivery resolution.
Produce hero loops and scroll sections for a launch page, with the product held steady by reference images.
Turn a key visual into a vertical motion piece for app stores, out-of-home screens, and social placements.
Direct short vertical spots with voiceover, effects, and music generated alongside the picture in one pass.
Move from a creative idea to a configured MiniMax H3 generation without leaving this model page.
Write the subject, action, environment, camera, style, timing, and sound. If you have reference media, upload it and explain the role of each asset.
Keep MiniMax H3 selected, choose the appropriate variant, mode, duration, aspect ratio, resolution, and audio settings, then check the displayed credit cost.
Start the task, follow progress in the result panel, review the completed video, reuse its settings for another take, or download the finished file.
MiniMax H3 is MiniMax's general-purpose omni-modal video model, released on 31 July 2026 and shipped in the Hailuo app as Hailuo 3.0. It understands text, images, video, and audio as one context and generates video with native stereo sound at up to 2K and up to fifteen seconds. It is the successor to Hailuo 2.3.
Most video stacks split the work across separate expert models: one for text-to-video, one for image-to-video, one for first-and-last-frame, others for subject reference, motion reference, and editing. H3 was pre-trained as a single model that takes the relationship between your references and your target shot as plain language, which is why one prompt can pull camera movement from a video, a character from an image, and a vocal from an audio clip.
Any whole number of seconds from 4 to 15, at 24 fps, in either native 2K or 768p. Supported aspect ratios run from 21:9 to 9:16, including 16:9, 4:3, 1:1, and 3:4. When you generate from a prompt alone you need to pick an aspect ratio explicitly; when you supply frames or references, the input decides the output shape.
Always. H3 returns 32 kHz stereo on every generation and there is no audio toggle to switch off, because picture and sound are predicted together rather than in two passes. Dialogue is stably supported in eleven languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.
Up to 9 images, 3 video clips, and 3 audio clips, with a ceiling of 12 files in total. Each video and audio clip must run between 2 and 15 seconds. Audio cannot be the only reference — it has to arrive with an image or a video so the model knows what the sound belongs to.
No, and the generator will tell you so before it spends credits. H3 exposes two input modes: a frame mode that takes zero, one, or two images as the opening and closing frames, and an omni-reference mode that takes the mixed image, video, and audio references. Frames and references belong to different modes, so pick the one that matches the control you need.
Credits are charged per second of output and depend on the resolution you pick, with 768p costing meaningfully less than 2K. The first five reference images are included; further images add a small per-image amount. The exact credit cost for your settings is shown in the generator before you start, and failed tasks are refunded automatically.
Yes. MiniMax published the complete weights on Hugging Face on 3 August 2026 under the MiniMax H3 Community License, including support for fine-tuning; the initial release ships full-attention inference, with the sparse-attention implementation to follow. Running them yourself takes serious hardware, so the generator on this page calls the hosted model instead.
Model capabilities and media on this page were researched from the developer's official product pages, announcements, and documentation.
Open the complete MiniMax H3 AI video generator above, add your prompt or references, and turn the next shot on your list into a finished video.
Start generatingBrowse all models