FastH3 Omni goes open: reference-to-video with sound on your own GPU, and the case for letting Opus cut cheap clips

On Oct 20, Hao AI Lab's FastVideo team released FastH3 Omni. It's a distilled MiniMax H3 built with Nuva Lab, and it comes with weights, inference recipes and training code. ComfyUI announced native support the same afternoon. What's new compared with the FastH3 builds we covered last week is the 'omni reference' input. Per the model card, you give it a prompt plus up to nine images, three videos and three audio clips (twelve in total, with at least one image or video). You point at them in the prompt with tags like <Picture 1>, <Video 1> and <Audio 1>. It renders 5, 10 or 15 seconds at 24 fps with synchronized 32 kHz stereo audio in eight sampling steps.

What it takes to run locally, from the FastVideo guide: an RTX 5090 (NVFP4 weights, 480p, text encoder offloaded), an RTX PRO 6000 (768p with everything resident), a DGX Spark (480p), or an RTX 4090 using FP8 weights streamed layer by layer, again at 480p. A Mac isn't supported yet because the MLX path doesn't handle references. In ComfyUI, drop the BF16 or INT8 model into diffusion_models, add the Qwen3-VL text encoder plus the video and audio VAEs, and use the MiniMax H3 Reference to Video node with 8 steps, guidance 1.0, video shift 12 and audio shift 3. Leave out the audio VAE and you get no sound. Expect big downloads: the encoder alone is about 25 GB.

The most useful take today came from @mickeyxfriedman. A 15-second 1080p Seedance 2.5 clip with references costs him about $11.31 through the API. FastH3 Omni on one rented B200 renders a 15-second 768p clip in roughly 75 seconds, about $0.14 a clip at ~$6.75 an hour, and Reactor's hosted API is about $0.18 a clip. He's upfront that a single FastH3 clip loses to Seedance side by side. His point is that Opus 5.5 in a coding agent closes most of the gap: it storyboards, keeps characters and voices consistent, reviews every take for artifacts, reshoots what drifts and cuts the 5–15 second clips into one scene. For his one-minute demo that came to 81 takes (496 seconds of footage) and $5–6 of GPU time. Even with that many reshoots, he estimates it's at least 20x cheaper.

Before you build on it: the code is Apache-2.0, but the weights fall under the MiniMax-H3 Community License. That licence requires written permission from MiniMax above $20M in yearly revenue and asks commercial products to display 'MiniMax H3' in their UI. Hosted throughput and the speedup claims come from the vendor and the creator, and we haven't reproduced them. Reference inputs make likeness and copyright your responsibility, so use footage, voices and characters you have rights to. Spotted, not tested. Credit Hao AI Lab, Nuva Lab and Mickey Friedman.