FastH3-Omni
MiniMax H3 omni reference, accelerated — generate video with synchronized audio from text, reference images, and keyframes.
Free preview · sign in with Google to generate
Generate a clip right here
Prompt in, video with synchronized audio out. Add reference images to keep characters on-model.
Result
Your clip will appear here
One model, every reference input
FastH3-Omni conditions a single clip on text, images, audio and keyframes. Video and audio generate together — nothing is synced in post.
Text prompt
Describe the shot with dialogue. H3 writes the audio into the same pass as the video.
Reference images
Up to 3 at realtime speed, up to 9 when it is not. Feed a character sheet or product shots so every clip starts from the same look.
Audio references
Up to 3 references — voices and soundtracks. The clip picks up timbre and rhythm from your audio.
First & last frame
Pin keyframe images where the clip starts and ends, so shots cut together at known frames.
Clip continuation
Latent mode carries motion, camera and sound into the next clip. Frame mode starts the next clip from the last frame.
Video + synced audio
About 5-15 seconds, 24 fps, in one generation — no dubbing, no separate sound pass.
From FastH3 to FastH3-Omni
FastH3 V1 and V2 made text-to-video fast. FastH3-Omni does the same for the omni reference mode.
| FastH3 V1 / V2 | FastH3-Omni | |
|---|---|---|
| Text to video with audio | Yes | Yes |
| Reference images | No | Up to 3 (realtime), up to 9 |
| Audio references | No | Up to 3 |
| First and last frame | No | Yes |
| Clip continuation | No | Yes |
What references let a team do
The same character in every clip
Give the character sheet or product shots as reference images. Each clip starts from the same look, so fewer clips fail review for drift.
A real voice and a real track
Give a voice line or a music clip as an audio reference. The video and the sound come from one generation, so you do not sync them later.
Shots that cut together
Pin the first or last frame to a storyboard frame or to the end of the previous shot. Ad versions and episodes join at known frames.
Longer scenes
Continue a clip in latent mode to keep motion, camera and sound across a scene that is longer than 15 seconds.
Audio references & continuation unlock Oct 20
How FastH3-Omni works
MiniMax H3 supplies the open-weight video-and-audio base model. FastVideo’s post-training makes it fast. Four ideas carry the whole pipeline.
Few-step post-training
A trained distilled path cuts the original H3 denoising schedule to four transformer calls — a fraction of the usual compute per clip.
Video Sparse Attention
The recommended Preview v1 profile applies sparse execution to eligible video-to-video attention tiles, skipping work that barely moves the output.
Optimized serving runtime
An inference runtime tuned for NVIDIA Blackwell runs FastH3 in production, with an optional NVFP4 checkpoint for higher throughput.
Video and audio, one pass
Every clip comes out with synchronized sound — the model writes picture and audio together, so nothing is dubbed or synced in post.
Published FastH3 results
Public FastH3 release results, not a universal claim — latency and quality still depend on the checkpoint, workload and hardware. Sources: FastVideo and Nuva Lab, October 2026.
Specifications
This site’s hosted lane currently generates at 480p / 768p in 16:9 and 9:16, 5 or 10 seconds. Full-resolution options follow the Oct 20 release.
The Oct 20 window
FastH3-Omni preview goes live with two exclusive launch partners.
FastVideo ships the final open weights with training code, DGX Spark and RTX support.
Audio references and clip continuation unlock on fasth3-omni.com as the new workflows land.
FastH3-Omni questions
What is FastH3-Omni?+
FastH3-Omni is a post-trained, accelerated version of MiniMax H3 with omni reference input. It takes text, up to three reference images (up to nine when realtime speed is not needed), up to three audio references, first and last frame keyframes and clip continuation, and outputs video with synchronized audio.
How is FastH3-Omni different from FastH3?+
FastH3 V1 and V2 cover text to video with audio. FastH3-Omni adds reference images, audio references, first and last frame keyframes and clip continuation — and keeps the same few-step acceleration.
When will FastH3-Omni be open?+
FastVideo has said the final FastH3-Omni weights ship on October 20, 2026, with training code and support for DGX Spark and RTX GPUs.
Can FastH3-Omni take a video as a reference?+
The published documentation lists text, image and audio references, keyframes and clip continuation. It does not list video references.
How fast is FastH3-Omni?+
The launch partner Reactor describes it as a realtime H3. For the earlier FastH3 V2, Hao AI Lab measured 36.5 seconds for a 5-second 1344×768 clip on one RTX PRO 6000.
Who offers the FastH3-Omni preview?+
Two exclusive preview launch partners: Nuva Lab — dedicated deployment, fine-tuning on your data, and the Creative Agent end to end — and Reactor, which runs the hosted realtime API. The open weights follow on October 20, 2026.
What is FastH3?+
FastH3 is FastVideo’s few-step, sparse-distilled derivative of MiniMax H3 for synchronized video-and-audio generation. The recommended Preview v1 profile uses four transformer calls and Video Sparse Attention.
Can FastH3 run locally?+
Yes. FastVideo has published local paths for NVIDIA DGX Spark and Apple Silicon, alongside data-center NVIDIA GPUs. Local does not necessarily mean realtime — latency and memory strategy vary materially by hardware.
Is FastH3 open source?+
The FastVideo framework is Apache-2.0 open source. The model weights are MiniMax H3 and derivatives, governed by the MiniMax H3 Community License — so "open weight" is the accurate description.
How are MiniMax H3, FastVideo and FastH3 related?+
MiniMax H3 is the base open-weight video-and-audio model. FastVideo is the Apache-2.0 post-training and inference framework. FastH3 is the resulting optimized H3 path — and FastH3-Omni extends it with reference images, audio references, keyframes and clip continuation.
Can US companies use FastH3?+
The MiniMax H3 Community License excludes the United States, European Union, United Kingdom and South Korea from its default territory; deployers there need MiniMax authorization. Nuva Lab holds an authorization for H3 deployments serving US companies.
Make your first clip
A prompt is enough to start. Reference images and keyframes keep every clip on-model.