MiniMax-H3 on TPU v6e
One prompt. Picture and sound, together.
A 33B omni-modal transformer that generates 768p video and 32 kHz stereo audio in a single pass — not a video model with a soundtrack bolted on. Drop-in compatible with the MiniMax API.
- Text to video10.1s
Starship bridge, jump to hyperspace
Cinematic, medium wide shot, pushing in slowly. In the cavernous, dimly lit bridge of a starship, a female captain stands silhouetted against a massive observation window. Outside, an armada of dreadnoughts begins to glow as thrusters charge. A blinding white flash floods the bridge as the fleet jumps.
- Resolution
- 1344×768
- Generated in
- 212s
- source
- Generated on our v6e-8
- First / last frame8.0s
Ramen, focus pulling to the family behind
A static shot of a patterned ceramic bowl of ramen in sharp foreground focus. Steam billows upward. The focus shifts deeper into the room and a family of seven comes into clarity, talking and eating, while the steam keeps rising in front of the lens.
- Resolution
- 1344×768
- source
- MiniMax reference sample
- Reference5.2s
Reference-driven edit, with a voice
An edit of a source video: a young man in a bright pink suit holds a black lamb in a sunlit pasture. His face is animated to speak new dialogue, with the voice timbre taken from a separate audio reference, over the original clip's background music.
- Resolution
- 1344×768
- source
- MiniMax reference sample
How it works
Three calls. The same shape you already use.
- 1
Submit
POST a prompt, a duration and an aspect ratio. You get a task id back immediately — generation takes minutes, so nothing blocks.
- 2
Poll
Query the task, or set a callback_url and we will POST you every status change, signed.
- 3
Collect
A finished task carries an expiring CDN link to an mp4 with an AAC stereo track.
curl https://api.h3.studio/v2/video_generation \
-H "Authorization: Bearer $H3_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "MiniMax-H3",
"content": [{ "type": "text", "text": "a lighthouse at dusk, waves on black rock" }],
"resolution": "768P",
"duration": 5,
"ratio": "16:9"
}'
{ "task_id": "424010985738629" }Built on measurements, not guesses
Every number on this site — latency, price, the shapes we can produce — comes from running the model on a v6e-8 and timing it. The tuning notes are open.
82 seconds, fastest tier
A 5-second clip, 49 denoiser steps, tensor-parallel across 8 chips. Slower tiers cost less per video and run more of them at once.
Compatible by construction
The request body matches platform.minimax.io field for field. Every deviation is a value restriction, and all of them are written down.
You are billed for what you asked for
Frame counts quantise upward, so a clip is often slightly longer than requested. You pay for the duration you sent.
Free credit, no card
Create an account and generate your first clip in a couple of minutes.