The Hao AI Lab released FastH3 Preview v1 on August 27, an open-weights model that generates synchronized video and audio in just four steps. It is a distilled version of MiniMax H3 that runs roughly 14 times faster than the base model while keeping most of the visual quality.

FastH3 ships through the open-source FastVideo project. Instead of the 49 transformer passes the base H3 needs per clip, the student model produces a full text-to-video-and-audio result in 4 passes, which is what collapses generation time from minutes to seconds.

What This Enables

If you already generate clips with H3, you can swap in the 4-step weights and iterate far faster. The team reports a 15-second clip on a single NVIDIA B200 dropping from 678.7 seconds on base H3 to 47.2 seconds with FastH3, and a 15-second 768p clip in under 13 seconds on an 8-GPU B200 setup. That turns a coffee-break render into a near-interactive loop, so you can test prompts, motion, and audio sync in one sitting. The weights are on the FastVideo model card, and the base model is the same one we covered in MiniMax H3 on fal.

Why It Matters for Creators

Few-step video models change the economics of short-form generation. When each attempt costs seconds instead of minutes, you can afford to reroll, which matters most for synchronized sound. H3's draw is that it generates picture and stereo audio together rather than dubbing after the fact, and it already has native ComfyUI workflows, the same pairing we walked through in the H3 sync-sound challenge. FastH3 keeps that joint output while making the retry loop cheap enough to actually use.

Key Details

Steps: 4 transformer passes, distilled from base H3's 49.

Method: Distribution Matching Distillation (DMD2) plus Video Sparse Attention at 90 percent sparsity, described in the DMD2 paper.

Speed: About 14x faster than base H3 on a single B200.

Output: Text to video and synchronized stereo audio in one pass.

License: MiniMax H3 Community License. Read the terms before commercial use.

What to Do Next

Clone FastVideo, pull the FastH3 Preview v1 weights, and run a few short prompts to compare quality against your current pipeline. Because this is a preview build governed by the H3 community license, check the usage terms before you put it into a client project, and keep the base H3 model available for shots where the extra steps buy you quality.