FastH3-Omni adds reference images and audio to MiniMax H3
The open-weight model previews on the Reactor API through October 20, when FastVideo says the final weights and training code ship.
Key takeaways
- FastVideo announced FastH3-Omni on October 13, 2026, an open-weight, reference-conditioned version of MiniMax H3 post-trained with Nuva Lab and Reactor.
- It generates video with synchronized audio from one prompt, in clips of about 5 to 15 seconds at 24 frames per second.
- It accepts up to 9 reference images (3 at realtime speed), up to 3 audio clips of 2 to 15 seconds, and first and last frame keyframes.
- FastVideo says the final FastH3-Omni weights and training code ship on October 20, 2026, with support for NVIDIA DGX Spark and RTX GPUs.
FastVideo's FastH3-Omni, a distilled version of MiniMax's realtime H3 that makes video and synchronized audio in one pass, entered an exclusive preview on Reactor's API on October 13, 2026, and leaves it on October 20.
The FastVideo team at Hao AI Lab announced FastH3-Omni that day, post-trained with Nuva Lab and Reactor, and Reactor opened the hosted preview under the model name reactor/fast-h3-omni. FastVideo says the final weights and training code ship on October 20, with support for NVIDIA DGX Spark and RTX GPUs.
The change is on the input side. Earlier FastH3 releases took text alone. FastH3-Omni is reference-conditioned: a prompt carries up to nine reference images, three of them at realtime speed, and up to three audio clips of 2 to 15 seconds each for a voice, a soundtrack or background sound. First- and last-frame keyframes pin the opening and closing shots, and a generated clip can be continued in latent or frame mode. Audio references need at least one image, so a supplied voice has a face to speak through.
Output is video with synchronized audio, about 5 to 15 seconds at 24 frames per second, on 16:9, 1:1, 9:16 and 4:3 canvases up to 1344x768. Reactor lists the speed as faster than realtime and bills the model at $0.75 a minute, or $0.0125 a second, down from a listed $2.10.
This is FastVideo's post-training work on MiniMax H3, MiniMax's open-weight base model. H3 makes video and audio in one pass but takes minutes per clip; FastH3 Preview v1, from August 27, cut the text-to-video-and-audio path to a four-step distilled one, and FastVideo later put FastH3 on DGX Spark and Apple Silicon through MLX. FastH3-Omni does the same distillation for the reference mode.
The open label carries a license caveat. FastVideo's framework is Apache-2.0, but the FastH3 weights stay under the MiniMax H3 Community License, which Nuva Lab says excludes the United States, European Union, United Kingdom and South Korea by default. Nuva Lab holds a MiniMax authorization for H3 deployments that serve US companies and offers the same model on dedicated hosts with fine-tuning on your own references.
FastH3-Omni previews on Reactor through October 20, 2026, at reactor.inc/models/fast-h3-omni; the open weights and training code follow the same day.
Sources
- docs.reactor.inc - Reactor docs: model name, clip length about 5-15s, 24 fps, aspect ratios, reference-image limits, audio references, keyframes, continuation
- reactor.inc - Reactor spec page: price $0.75/minute ($0.0125/sec), faster-than-realtime, up to 9 pictures, up to 3 audio clips, discrete clip lengths
- reactor.inc - Reactor model page and price markdown from $2.10/minute
- x.com - Reactor launch post, Oct 13 2026: exclusive preview until 10.20.2026
- x.com - FastVideo/Hao AI Lab announcement: post-trained with Nuva Lab and Reactor, open weights + training code Oct 20, DGX Spark and RTX
- nuvalab.ai - Nuva Lab: inputs table, output specs, two exclusive preview partners, licence territory, Oct 20 weights
- nuvalab.ai - Nuva Lab launch post, Oct 14 2026: FastH3 vs FastH3-Omni table, inputs
- haoailab.com - FastH3 Preview v1 (Aug 27 2026): base H3 takes minutes, four-step distillation, T2VA-only prior release
- haoailab.com - FastH3 on DGX Spark and Apple Silicon via MLX
- huggingface.co - MiniMax H3 base open-weight model page