MiniMax H3 is a general-purpose multimodal generation model: it reads text, images, video and audio as one context and returns video with native stereo sound, up to 15 seconds at 2K. MiniMax positions it on price-performance, quoting under a third of mainstream per-second cost at 2K, and has stated it plans to open the model weights.
Capability snapshot
2K
default resolution
15s
maximum clip
stereo
native audio
Practical value mapped to real production workflows.
Benefit
01
One context, not four pipelines
Text, image, video and audio go in as a single context, so a prompt can say "take the camera move from this clip, have the person in this image sing, match this voice" and be understood as one instruction rather than routed through separate expert models.
Benefit
02
Sound generated with the picture
Audio is produced jointly with the video and comes out as native stereo, so dialogue, effects and ambience arrive already aligned to the shot instead of being layered on afterwards.
Benefit
03
2K at a lower per-second cost
2K is the default rather than a premium tier. MiniMax quotes H3's per-second price at 2K as under a third of mainstream models, and under half of mainstream 720p when run at 768p — which changes what volume iteration costs.
A simple three-step path from source material to a reusable result.
Step 1
Assemble the context
Bring the reference material you have — a clip for camera movement, an image for the subject, audio for the voice.
Step 2
Describe the relationship
State how the references relate to the target video in plain words. H3 resolves the cross-modal reasoning itself.
Step 3
Generate and iterate
Render up to 15 seconds at 2K with stereo audio, then iterate — the per-second cost is what makes repeated passes practical.
Up to 15 seconds, with 2K offered as the default resolution. A 768p tier is also available and is where MiniMax quotes its lowest per-second pricing.
H3 generates audio jointly with the video, and the output is native stereo rather than a mono track mixed in afterwards.
MiniMax announced an intention to open the model weights, citing hardware compatibility and customization as design goals. Treat availability as pending until the release lands rather than as a current guarantee.
Next step
Open the tool with a real image, compare the results, and keep the version that works.