JUL 30, 2026
MiniMax H3 vs Seedance 2.0: 2K with native audio, three ways to drive it
July 30, 2026
Table of contents
- Text to video
- A start frame, with speech
- Omni references — image, video, and audio together
- What it costs
- Where H3 fits
MiniMax’s newest video model, H3, is live on POST videos/create as model: "Hailuo-3.0". Two things separate it from everything else in the MiniMax lineup: every generation ships with native stereo audio — no separate voice or music pass — and it renders 2K at 24 fps by default, with the short edge at 1440 pixels (2560×1440 at 16:9, 1440×2560 at 9:16).
H3 is a general-purpose multimodal model rather than a set of task-specific modes. The same endpoint takes a prompt alone, a start frame, or a tray of image, video, and audio references, and the request shape is the one Seedance 2.0 already uses — set model, resolution, and duration directly.
The three clips below are the actual output of a single 4-second run at each mode, unedited. Turn the sound on — the audio in all three is generated by the model.
The last of the three is a direct comparison: it reuses the exact image, video, and audio references and the exact prompt from our Seedance 2.0 omni post, so H3 and the three Seedance 2.0 tiers can be judged on identical input rather than on separate demos.
Text to video
No references at all, just a prompt. aspectRatio is explicit here, though H3 defaults to 16:9 when no reference is supplied.
Aerial push over jagged sea cliffs at dawn. Waves crash white against black rock far below as mist burns off in the first gold light. A lone seabird crosses frame. Cinematic wide-angle, natural color grade. Crashing surf, wind, and distant gull calls.
curl — Hailuo-3.0, 2K, text-to-video, 16:9
curl --location 'https://api.useapi.net/v1/minimax/videos/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="Hailuo-3.0"' \
--form 'prompt="Aerial push over jagged sea cliffs at dawn. Waves crash white against black rock far below as mist burns off in the first gold light. A lone seabird crosses frame. Cinematic wide-angle, natural color grade. Crashing surf, wind, and distant gull calls."' \
--form 'resolution="1440"' \
--form 'duration="4"' \
--form 'aspectRatio="16:9"'
The surf, wind, and gulls are all model-generated — nothing was added to the audio track.
A start frame, with speech
Supply one image as the first frame and let the prompt drive what happens next. The portrait below is the same reference used in our Seedance 2.0 omni post, so the two models can be compared on identical input.
fileID — the start frame. Upload it with POST /files first.
She looks into the camera, smiles, and says warmly: “You made it — I was starting to think you’d stood me up.” Warm bar lights bokeh behind her, subtle handheld drift on the camera. Low jazz and crowd murmur sit under her voice.
curl — Hailuo-3.0, 2K, start frame + speech, 9:16
curl --location 'https://api.useapi.net/v1/minimax/videos/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="Hailuo-3.0"' \
--form 'prompt="She looks into the camera, smiles, and says warmly: '\''You made it — I was starting to think you'\''d stood me up.'\'' Warm bar lights bokeh behind her, subtle handheld drift on the camera. Low jazz and crowd murmur sit under her voice."' \
--form 'fileID="user:12345-minimax:1234567890-file:your-image"' \
--form 'resolution="1440"' \
--form 'duration="4"' \
--form 'aspectRatio="9:16"'
Spell the spoken line out in the prompt. H3 renders the delivery from the text, so quoting the line and saying it is spoken is what puts a voice on the track.
Omni references — image, video, and audio together
The same three references from the Seedance 2.0 omni post, and the same prompt, run through H3 for a direct comparison. H3 accepts up to 9 images, 3 videos, and 3 audio clips, capped at 12 files in total, each addressed by an @-tag in the prompt.
@video1 — the scene, rope, and motion to keep. Its face is blurred on purpose: a clear real face in a video reference is rejected by the content filter, and the identity comes from @image1 anyway.
@audio1 — the speech to drive, about 2 seconds. Together with the portrait above as @image1, that is the full tray.
Replace the woman in @video1 with @image1, keeping the same outdoor scene, rope, and jumping movements. She speaks out loud, clearly saying “Count me jumping” in the voice from @audio1, audible throughout the clip.
curl — Hailuo-3.0, 2K, omni image + video + audio
curl --location 'https://api.useapi.net/v1/minimax/videos/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="Hailuo-3.0"' \
--form 'prompt="Replace the woman in @video1 with @image1, keeping the same outdoor scene, rope, and jumping movements. She speaks out loud, clearly saying '\''Count me jumping'\'' in the voice from @audio1, audible throughout the clip."' \
--form 'fileID="user:12345-minimax:1234567890-file:your-image"' \
--form 'videoFileID1="user:12345-minimax:1234567890-file:your-video"' \
--form 'audioFileID1="user:12345-minimax:1234567890-file:your-audio"' \
--form 'resolution="1440"' \
--form 'duration="4"'
No aspectRatio was sent here. With references attached H3 defaults to Auto and matched the 9:16 shape of the inputs on its own.
What it costs
H3 bills 12 credits per second of output at 2K. A video reference is charged on its own length, at the same rate, on top of the generated clip — image and audio references add nothing per second.
| Request | Credits |
|---|---|
| 4-second clip, no references | 48 |
| 4-second clip, image reference | 48 |
| 4-second clip, 10-second video reference | 168 |
| 15-second clip, no references | 180 |
That surcharge is worth planning around: trim a video reference to the shortest clip that carries the motion you need, since its full length is billed whatever the output duration.
Where H3 fits
Seedance 2.0 still reaches 4K and runs cheaper per second at 480p and 720p, so it remains the better pick for volume iteration and for the highest resolution. H3 is the one to choose when you want 2K with sound in a single call — the native audio track removes an entire step from the pipeline, and 24 fps at 1440 is a materially different starting point for finishing work.
See the POST videos/create reference for every parameter, the reference-media limits, and a live Try-It console.