MiniMax FastH3 V2

Overview

FastH3 V2 is FastVideo's eight-step distilled version of MiniMax H3 for text-to-audio-video generation. From one written prompt, it creates a landscape video with matching visuals and synchronized audio, including ambience, sound effects, music, and spoken dialogue.

V2 re-distils the diffusion transformer only. Compared with the four-forward Preview v1 checkpoint it uses eight transformer forward passes instead of four, a nine-point sigma grid instead of five, and a denser 80 percent sparse video attention policy instead of 90 percent.

What you can build

• Short cinematic scenes with spoken dialogue
• Product and campaign videos with matching sound design
• Character moments with environmental ambience and voice
• Social clips, concept shots, trailers, and visual story beats
• Nature, city, action, and atmosphere footage with synchronized effects

Inputs

• Prompt: Describe the subject, action, setting, camera movement, lighting, visual style, ambience, sound effects, and dialogue. Put the exact words to be spoken in quotation marks and identify the speaker.
• Duration: Choose a whole-number duration from 5 to 15 seconds. FastH3 uses a causal video frame grid, so the generated clip length is the nearest supported native duration rather than an exact wall-clock match.
• Resolution: 768P is the native canvas tier and gives the most detail. 480P is available for lighter output.
• Seed: Optionally set a seed from 0 to 4294967295 for repeatable iterations.

Outputs

The model returns one MP4 containing both video and audio.

• Resolution: 1344 x 768 landscape at 768P
• Frame rate: 24 FPS
• Video codec: H.264
• Audio codec: AAC
• Audio can include synchronized speech, ambience, music, and effects requested in the prompt

Recommended settings

• Use 768P for final output; it is the canvas the checkpoint was trained on.
• Start with a 5- or 6-second clip while refining composition and motion.
• Describe one clear primary action and keep the camera direction explicit.
• For speech, name the speaker, put the line in quotation marks, and specify language, tone, and delivery.
• Describe background sound separately from dialogue so both are represented.
• Keep the same seed when comparing prompt changes, then change it to explore a new composition.

Limitations

• This checkpoint supports text-to-audio-video only. It does not support image-to-video, first-frame conditioning, or reference-video input. FastVideo states that FL2VA and Ref2VA were not distilled.
• Duration choices are mapped to supported native frame counts. The 15-second option maps to the checkpoint maximum of 345 frames, or approximately 14.375 seconds at 24 FPS.
• Lip synchronization, exact wording, character identity, fine text, and complex physical interactions can vary between generations.
• Detail at 480P is lower than at the native 768P tier.

Model and license notes

FastH3 V2 is published by FastVideo and is based on MiniMax H3. Review the checkpoint's model card and license terms before using generated media in production or commercial workflows.