FastH3-Omni preview speeds up MiniMax H3 reference video
FastH3-Omni is an accelerated version of MiniMax H3 that conditions clips on images, audio, text and first and last frames. A hosted preview is live, with open weights and training code announced for October 20, 2026.
What it does
FastH3-Omni is a sped-up version of MiniMax H3's reference mode, post-trained by FastVideo with Nuva Lab and Reactor. A clip can be guided by a text prompt, up to 3 reference images at realtime speed (up to 9 when realtime isn't needed), up to 3 audio references, and first and last keyframes.
It outputs about 5 to 15 seconds of video with synced audio at 24 fps, and can continue a clip to build longer scenes.
Results and limits
Reactor calls it a realtime H3, but no measured time per clip is public yet; Nuva Lab says it will add numbers after the open release. For comparison, the earlier text-to-video FastH3 V2 took 36.5 seconds for a 5-second 1344×768 clip on one RTX PRO 6000.
The source notes that the October 20 open release may change these limits.
Worth knowing
It's available now as a preview through Nuva Lab dedicated hosts and the Reactor API (until October 20). Open weights and training code are scheduled for October 20, 2026. A keyframe needs at least one normal reference image in the same request.
Try it yourself
Give [character sheet image] as reference, [voice line audio] as audio reference, and [storyboard frame] as the first frame. Prompt: "[Character] [action], says '[line]'." Then continue the clip in latent mode for the next beat.