Videos · Models
MiniMax-H3
Open-weight video and audio generation from text, frames, or references. One request returns a synchronized MP4.
Specified
Overview
| Property | Value |
|---|---|
| Model name | minimax/h3 |
| Tasks | t2va, fl2va, ref2va |
| Output | H.264 MP4 at 24 fps, stereo AAC audio at 32 kHz |
| Length | 4 to 15 seconds |
| Resolution | 768p on the short edge, for example 1344×768 at 16:9 |
| Outputs per job | 1 to 10 |
| Quality | lossless and turbo at launch |
| Typical generation | 50 to 150 seconds at full quality; turbo is several times faster |
Behavior to design around
- 768p is the ceiling. A 2K mode exists only in the vendor's own hosted service, not in the open checkpoint we run.
- Two checkpoint pools.
t2vaandfl2vashare one;ref2vahas its own. Each pool queues independently. - 15 seconds per job. Longer clips need chained jobs; see Tasks and inputs.
Request fields by task
Every request needs model and task. Beyond that:
| Field | t2va | fl2va | ref2va |
|---|---|---|---|
prompt | required | optional | optional |
conditions | not used | 1 or 2: first_frame / last_frame with image_url | 1 or more: image / video / audio with url |
seconds | 4 to 15 | 4 to 15 | 4 to 15 |
target | short_edge: 768 | short_edge: 768 | short_edge: 768 |
quality | lossless, turbo | lossless, turbo | lossless, turbo |
num_outputs_per_prompt | 1 to 10 | 1 to 10 | 1 to 10 |
seed | optional | optional | optional |
Full request and response shapes are in the Videos API reference.