Updated
What is MiniMax H3?
MiniMax H3, released July 31, 2026, is MiniMax's flagship video model. It renders 768p or 2K video at 24 fps with native audio, runs 4 to 15 seconds per generation, and takes four input styles: text alone, a start frame with an optional end frame, 1 to 9 reference images, or up to 3 reference videos. For drafts and high volume, MiniMax H3 Lite takes the same inputs at 480p or 768p for about half the per-second price; H3 Fast adds reference audio and a seed in about 30 seconds a job, H3 Max carries the full control set at 480p or 768p, and H3 Max Turbo is its distilled text-or-frames version. Resolution and aspect ratio are selected independently for text and reference generation; start/end-frame generation follows the supplied frame shape.
Key features of MiniMax H3
Motion physics at 2K
House test at 2560x1440, 8 seconds, prompt: slow-motion water pouring over glass on dark slate. The churn inside the glass, the refraction through it, and the droplets beading on the stone all track convincingly. The model simplified the brief from a stack of glasses to one, and the pour briefly smears into a sheet before the splash resolves.
Native audio in the same pass
House test at 2560x1440, 8 seconds: a violinist under a stone archway at dusk. The file arrives with an audio track generated alongside the picture, no separate audio step. Framing, hand placement on the bow, and the warm lantern light against the blue-hour street all hold as the camera circles.
21:9 cinematic output
House test at 2944x1248, 10 seconds: a dawn dolly down a rain-soaked neon street. Sign text renders legible and its color spills into the wet-asphalt reflections the way real signage does. The steam plumes rise a little too symmetrically on both sides of the frame.
Reference-to-video
Pass 1 to 9 reference images and name them in the prompt as @Image1 through @Image9. The model keeps the subject consistent across generations and can combine a character, a product and a location in one shot. Reference images cannot be combined with start/end frames in one request.
Reference videos
Up to 3 reference videos, 15 seconds in total, passed in video_urls. Each second of uploaded video is billed at the 2K per-second rate, whichever output resolution you choose.
Start and end frame control
A start frame on its own, or a start and an end frame together. With a frame image set, output follows the input frame instead of aspect_ratio.
What MiniMax Hailuo H3 is best for
Character-locked series
Subject reference images keep the same character across every video in a campaign.
Character-in-location shots
Pass the character as @Image1 and the setting as @Image2, and the prompt places one inside the other.
Cinematic 21:9 spots
The 2944x1248 widescreen tier renders trailer-shaped output straight from a text prompt.
Sound-on social video
Native audio means output is publishable without an audio pass.
Start-and-end frame ads
Pin the open and the close as key art and let the model animate the move between them.
MiniMax Hailuo H3 use cases
Run a product campaign where the same mascot appears in ten videos by sending the character sheet as reference images. Put a product into a new scene by passing the product and the location as two reference images. Render a 21:9 teaser at 2944x1248 for a landing-page hero. Build a vertical 9:16 spot with sound for social feeds in one call. For pinned compositions, pass a start frame, optionally with an end frame, and the output follows the input frame's ratio.
MiniMax Hailuo H3 limitations
MiniMax H3 caps at 15 seconds per generation and 24 fps, with 768p and 2K output but no 4K tier; rivals like Kling 3.0 render native 4K. Reference images and frame images are mutually exclusive in one request. The UnificAlly route takes reference images, up to 3 reference videos and up to 3 reference audio clips (videos and audio 15 seconds in total), plus a seed and a prompt expansion level. On the Artificial Analysis generation boards it sits behind Gemini Omni Flash on Text-to-Video and behind Gemini Omni Flash and Seedance 2.0 on Image-to-Video with audio.
MiniMax Hailuo H3 API examples
Call MiniMax H3 from any language by POSTing to /v1/tasks. Full parameter docs live at docs.unifically.com/models/video/minimax/hailuo.
curl -X POST https://api.unifically.com/v1/tasks \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "hailuo/minimax-h3",
"input": {
"prompt": "A slow dolly shot across a rain-soaked neon street, reflections rippling in the puddles",
"aspect_ratio": "21:9",
"resolution": "2K",
"duration": 10
}
}'
Successful submission returns a task_id. Poll GET /v1/tasks/<task_id> or set a callback_url on the request to receive the finished result.
FAQs
People also ask about MiniMax Hailuo H3
MiniMax H3, released July 31, 2026, adds selectable 768p and 2K output at 24 fps, stretches duration to any length from 4 to 15 seconds, adds 21:9 and three more aspect ratios, generates native audio in the same pass, and adds prompt-only text-to-video plus omni-reference generation with up to 9 images. Hailuo 2.3 on UnificAlly takes a start frame only.
MiniMax H3 supports 768p and 2K output through the resolution parameter. The default is 2K; aspect_ratio independently controls the output shape when no start or end frame is supplied.
No. image_urls cannot be combined with start_frame_url or end_frame_url in one request. Use frames for pinned compositions and reference images for subject or style control. An end frame also needs a start frame.
Yes. Pass up to 3 reference videos in video_urls, 15 seconds in total. Each second of uploaded video is billed at the 2K per-second rate ($0.117/sec), whichever output resolution you choose.
Yes. MiniMax H3 Lite takes the same inputs and outputs 480p or 768p at about half the per-second price. H3 Fast, H3 Max and H3 Max Turbo also output 480p or 768p; Fast is the quick-turnaround one, Max keeps the full control set, Max Turbo is distilled and takes text or frames only. Use H3 when you need 2K.
Yes. Audio generates in the same pass as the video, so output arrives with sound and no separate audio step.
On the Artificial Analysis boards (July 31, 2026), MiniMax H3 is #1 in Video Editing at Elo 1,130, #2 in Text-to-Video with audio at Elo 1,242, and #3 in Image-to-Video with audio at Elo 1,185. Video editing is not exposed on the UnificAlly route, which generates new video from text, frames or reference images.
Yes. MiniMax H3 requires a prompt of 1 to 2000 characters on every request, including frame and reference runs.
Any whole second from 4 to 15 per generation. The default is 5.
