FastH3 Omni: Open 8-Step Reference-to-Video With Audio, From DGX Spark to B200

The FastVideo project has released FastH3 Omni, an open video-generation model that turns a text prompt plus a stack of image, video and audio references into video with synchronized stereo audio, in just eight transformer evaluations. It is a distilled, reference-aware variant of MiniMax-H3, built with Nuva Lab, with compute from Reactor and help from NVIDIA's FastGen team, and it runs on hardware from B200 servers down to an RTX 5090 or DGX Spark.

Weights and training code are public, but the weights sit under the MiniMax-H3 Community License, not a conventional open-source license. This post covers what the model does, how to run it, the speed claim, and the license question. It builds on our earlier coverage of FastVideo's first FastH3 release for real-time open video. Source: the FastH3 Omni model card.

TL;DR: FastH3 Omni at a glance

What is new compared with the first FastH3?

The first FastH3 releases distilled MiniMax-H3 into few-step text-to-video models. Omni adds reference conditioning: you can feed ordered references and refer to them in the prompt, so characters, objects, styles and sounds from your own images, clips and audio carry into the output. Video references bring their soundtracks with them. That puts it in the same product category as the reference features in commercial models, such as the Omni Reference tool in Kling 4.0 that we covered in our Kling 4.0 launch post, but with weights you can run yourself.

The card describes the release as the "final BF16 release checkpoint" and points to a family of variants: FP8, NVFP4 with quantization-aware distillation (QAD), ComfyUI single-file versions, optional NVFP4 VAE decoders in a 36-layer quality tier and an 8-layer fast tier, and an MLX build for Apple silicon that appeared the day before.

How do you run it?

The model card routes you to a runtime by hardware.

Details that matter in practice:

• Download the BF16 checkpoint with the Hugging Face CLI. The repository contains the reference transformer, schedulers and inference configuration; the runtime loads the shared text encoder, tokenizer, processor and video and audio VAEs from MiniMax-H3, so you must prepare those components too.
• Local profiles default to 480p. The Spark profile keeps model components resident; the RTX 5090 offloads the text encoder and needs a precomputed AdaLN table.
• NVFP4 needs Blackwell. The bundled NVFP4 text encoder requires it. Hardware without FP4 support needs compatible BF16 encoder components.
• Quality differs by precision. The card warns that FP8 and NVFP4 QAD outputs can differ from BF16.

If you are choosing local hardware for this kind of workload, our comparisons of NVIDIA's RTX Spark power envelopes and the Gigabyte deskside GB300 against DGX Spark show where each fits.

Why only eight steps?

Diffusion video models normally refine a noisy latent over dozens of steps, each a full pass through a very large transformer. Distillation trains a student that reaches a similar result in a handful of steps. FastVideo's earlier releases used distribution-matching distillation (the Hub tags cite DMD2) to cut MiniMax-H3 to four and eight steps, and the Omni version applies eight evaluations with guidance fixed at 1.0, which also removes the extra pass that classifier-free guidance normally needs. Fewer steps and no guidance pass are the two main reasons the model can run on a single consumer card at all.

The cost is fidelity: distilled models can lose fine detail or diversity compared with the full-step teacher. The card's own caution about precision differences between BF16, FP8 and NVFP4 applies on top of that. Expect to trade some quality for speed and plan a side-by-side comparison with the base model if quality is critical.

A rocket on a launch platform, illustrating how FastH3 Omni ships an 8-step video model for server and local runtimes

A sensible first run

1. Pick the runtime for your hardware from the table above, and read the matching guide before downloading checkpoints. The BF16 weights alone are large, and the shared MiniMax-H3 components must be fetched separately.
2. Start with the 480p local recipe and a single image reference, so you can confirm the pipeline produces audio and video before adding more references.
3. Add references one at a time. The card says references are ordered and tagged in the prompt, and video references carry their soundtracks, so a clip with music will influence the audio.
4. Compare the NVFP4 and FP8 variants against BF16 on the same prompt and seed, and keep the one that meets your quality bar at an acceptable speed.
5. If you use ComfyUI, start from the Comfy single-file variants, which the card lists for visual workflows.

The "12x faster, real time" claim

Posts from the FastVideo team, relayed by news aggregators, say FastH3 Omni reaches real-time generation roughly 12x faster on NVIDIA Blackwell. Early users on social media have compared it favorably with fal.ai's H3 Max, citing the high quality and the fully open pipeline, including training code. We did not find the benchmark setup, resolution or clip length behind the 12x figure on the model card, and the comparison against H3 Max is a user's opinion. Reproduce on your own hardware at the resolution you need before planning around it. Real-time claims typically apply to short clips at low resolution on top-end GPUs, and the local profiles here default to 480p.

What can you build with reference-to-video?

Based on the input design, plausible uses include character-consistent shorts built from a handful of reference images, music-driven clips using audio references, and style transfer from a reference video with its soundtrack. The card itself documents inputs, not showcase results, so test your own prompts. If you work with video editing pipelines, see our posts on Mirelo SFX 2.0 for adding or removing sound and fal's Flux 3 Max video pricing for the commercial alternatives.

The license question

The card says use of the weights "is governed by the MiniMax-H3 Community License Agreement," linked from MiniMax-H3's repository, and the Hub tags list the license as "other." In practice, that means:

1. It is not Apache 2.0 or MIT. Read the community license for restrictions on commercial use, scale thresholds, output use and redistribution.
2. Derived quantizations and ComfyUI conversions inherit the base terms.
3. The surrounding code, training recipes and FastVideo project are open, but the model weights are the licensed asset.

Acknowledged collaborators include Nuva Lab, which co-developed the model, defined the omni-reference dataset and offers dedicated production deployment, fine-tuning and a creative agent on top. That is a commercial offering, and the card says so openly.

Where it sits in the open video landscape

Open video generation has moved quickly from silent, short clips to models that produce audio and accept multiple references. FastH3 Omni is notable less for a new base model than for engineering: it shows that a large commercial-scale video model can be distilled, quantized to four-bit floating point and fitted on a 32 GB gaming card, with a documented path from a single desktop GPU to a rack of B200s. That makes it a useful reference for anyone deciding whether to self-host video generation or rely on APIs. For a view of the hosted market and its pricing, compare our coverage of Higgsfield's unlimited Seedance 2.5 plan.

Limits and what to check

• Resolution and length depend on the runtime and memory; the card does not give a table of supported sizes.
• Quality versus precision. Compare BF16, FP8 and NVFP4 outputs on your own prompts.
• Content and consent. Reference-based generation can reproduce a person's likeness from images, video or voice clips. Only use references you have rights to, and consider disclosure norms for synthetic media.
• Independent benchmarks. None yet that we found; community ComfyUI and GGUF conversions are appearing quickly, so expect user tests this week.

Related reading

• FastH3 v1: real-time open video on Blackwell
• Kling 4.0 launch
• Video generation guide: Sora, Runway, Kling
• Perplexity Portable Computer with DGX Spark
• NVIDIA RTX Spark power envelopes
• Mirelo SFX 2.0
• Official: FastH3 Omni model card, FastVideo on GitHub and UniServe

Details are accurate as of October 22, 2026 and come from the Hugging Face model card. Speed claims are the project's own.