MiniMax FastH3 V2

Overview

FastH3 V2 is an eight-step FastH3 checkpoint from FastVideo, distilled from MiniMax H3, MiniMax's open-weight audio-video model. From one text prompt it generates the picture and a stereo soundtrack together: dialogue, ambience, foley and music.

V2 retrains only the diffusion transformer, again with data-free DMD2 on prompts alone. The text encoder and the video and audio VAEs are the same files as in Preview v1. It runs eight transformer passes on a nine-point sigma schedule, where Preview v1 runs four on five and base H3 runs 49. Its video sparse attention is also denser, at 80 percent sparsity instead of 90.

Choosing a variant: MiniMax FastH3 is the four-pass Preview v1 checkpoint and the lighter of the two, with the same inputs and output format. V2 spends twice as many transformer passes on each clip. To see what that changes on your material, run one prompt and seed on both.

What you can build

• Nature shots with matching ambience, like the fox in a frosted field and the floating lanterns at dusk on this page
• Craft and performance clips where the sound comes from the action: a blacksmith at the forge, a violinist on a subway platform
• Workplace scenes such as a cockpit or a restaurant kitchen for explainers and training videos
• Vertical 9:16 social clips, square 1:1 ad slots and 21:9 trailer openers

Inputs

• Prompt: describe the subject, action, setting, camera and light, then the sound. Put spoken words in quotation marks and name the speaker.
• Duration: 5 to 15 whole seconds, snapped to the nearest length H3 supports at 24 FPS. 5 gives 5.17 seconds (124 frames) and 15 gives 14.375 seconds (345 frames).
• Resolution: 480P (the default) or 768P, H3's native tier.
• Aspect ratio: 16:9 (default), 9:16, 1:1, 4:3, 3:4 or 21:9.
• Seed: optional, from 0 to 4294967295, 1000 by default.

Outputs

One MP4 with H.264 video at 24 FPS and AAC stereo audio at 32 kHz. At 768P, 16:9 is 1344 x 768, 9:16 is 768 x 1344 and 21:9 is 1536 x 672. At 480P the short side is 480 pixels. The six samples here are 768P 16:9 at the 5-second setting.

Recommended settings

• Render finished clips at 768P.
• Keep one main action per clip and name the camera move.
• For dialogue, give the speaker, the quoted line, the language and the delivery, and describe background sound in its own sentence.
• Hold the seed fixed while you edit the prompt.

Limitations

• Text-to-audio-video only. FastVideo did not distill H3's first/last-frame (FL2VA) or reference (Ref2VA) modes, so there is no image or video input.
• FastVideo notes that difficult motion, fine detail and some audio may fall short of base MiniMax H3.
• Output stops at 768P and 14.375 seconds. MiniMax's separate 2K stage is not part of this model.
• Lip sync, exact wording, character count and on-screen text change from seed to seed.

Safety & compliance

The weights are released under the MiniMax H3 Community License, inherited from the base model, and the bundled Qwen3-VL-32B text encoder is under Apache 2.0. The license does not cover use in the United States, the European Union, the United Kingdom or South Korea, and that exclusion applies to generated clips as well. It also requires clips posted publicly to be clearly marked as machine-generated and bars using them to improve other AI models. Read the full terms before commercial use, and do not depict real, identifiable people or their voices without consent.