MiniMax-H3
joint video + audio · 8× Ascend 910B3 · Atlas 800I A2
checking…
Access code
Mode
T2VA — text to video + audio
FL2VA — first/last keyframe to video + audio
Prompt
insert structured template
H3 was tuned on
structured
prompts with three labelled fields, not plain sentences. The official pipeline's rewriting stage isn't in the open release, so structure it yourself for a large quality gain — see
docs/09-hyperparameters.md
.
First frame
Last frame
FL2VA needs at least one keyframe. Supply both to interpolate between them.
Duration (s)
Inference steps
Advanced
Width
Height
Seed
flow_shift
audio_flow_shift
1344×768 is the upstream-validated 768p shape on NPU. Other shapes are untested here.
Generate
One request uses all 8 NPUs, so jobs run strictly one at a time. A 5 s clip takes roughly 2–3 minutes.