FastH3

FastVideo FastH3 Preview v1

FastH3

Open-weight H3 generation in four transformer forwards.

FastH3 is FastVideo's step-distilled MiniMax H3 preview for text-to-video-and-audio generation. This independent guide brings together the recommended checkpoint, verified commands, official samples, and the limits you need to know before running it.

View recommended model Run FastH3 locally

FastH3 at a glance

4

Transformer forwards

The recommended FastH3 Preview v1 checkpoint follows a trained five-point sigma grid that results in exactly four DiT calls.

T2VA

Synchronized output

The current preview generates video and stereo audio together from a text prompt in one pipeline.

14x

Reported peak speedup

FastVideo reports up to 14x speedup on a single NVIDIA Blackwell GPU under its published benchmark setup.

90%

Trained VSA sparsity

The recommended VSA / Data-Free checkpoint was trained with VSA-H3 at 90% sparsity using tile size 64.

Model overview

What is FastH3?

FastH3 is not a new base model and it is not the same product as the hosted MiniMax H3 app. It is a FastVideo checkpoint family distilled from the open MiniMax H3 base weights to reduce the number of denoising evaluations while preserving synchronized video and audio generation.

01 / DMD2

Four-step DMD2 distillation

The original MiniMax H3 base sampler uses 50 denoising steps. FastH3 Preview v1 compresses generation into four transformer forwards using DMD2 distillation. Fewer evaluations target lower latency, while the preview status signals that difficult motion and fine detail may still trail the base model.

02 / VSA-H3

Sparse attention built into training

The recommended checkpoint was trained with FastVideo's VSA-H3 attention path, 90% sparsity, and 64-token tiles. That sparse route is part of the model's operating point. Removing VSA or treating dense attention as a drop-in replacement changes the intended setup.

03 / T2VA

Video and audio in one pipeline

FastH3 Preview v1 currently supports text-to-audio-video generation. It reuses the MiniMax H3 text encoder, video VAE, audio VAE, tokenizer, and scheduler, while replacing the base transformer with the distilled student or applying the matching LoRA.

04 / WEIGHTS

Open weights with a community license

FastVideo publishes full checkpoint weights and a pre-extracted LoRA. FastH3 inherits the MiniMax H3 Community License, so developers should review its territory, acceptable-use, and redistribution terms before production or commercial use.

Validation gallery

Watch real FastH3 outputs

These clips come from the FastVideo team's public Preview v1 validation gallery. They include generated audio. Use the player controls to hear the synchronized output and inspect the actual checkpoint rather than a staged product mockup.

208 frames

Paper-craft tutorial

A multi-shot instructional scene with Japanese title cards, paper-folding actions, close foley, and no background music. This sample tests readable structure, controlled camera movement, and sound synchronized to hand actions.

Prompt focus: paper-craft instructor, overhead camera, Japanese labels, crisp fold sounds, quiet studio ambience.

259 frames

Anime action sequence

A four-shot fantasy action scene with two sky-mages, Chinese dialogue, rapid camera changes, energy effects, impacts, wind, and environmental audio. It exercises motion, shot continuity, dialogue, and layered sound.

Prompt focus: cel-shaded sky-mages, four timed shots, Chinese dialogue, dynamic camera motion, synchronized impacts.

124 frames

Cinematic portrait

A restrained live-action close-up in a dim room with subtle breathing, dress movement, a ticking clock, and sparse piano. It shows how FastH3 handles low-motion performance and a deliberately quiet soundscape.

Prompt focus: static medium close-up, dim practical light, subtle expression, room tone, clock ticks, sparse piano.

Sample descriptions and technical metadata are summarized from the official FastH3 Preview v1 release gallery. Results vary by prompt, seed, hardware, runtime, and checkpoint.

Model selection

Choose the right FastH3 checkpoint

FastVideo's Preview v1 collection contains four full-weight checkpoints plus a matching LoRA repository. Start with VSA / Data-Free unless you are studying an ablation or need the adapter-only download.

VSA / Data-Free

Recommended

The main FastH3 Preview v1 release. It uses prompt-only DMD2 training, VSA-H3 at 90% sparsity, tile size 64, and four DiT calls. The official validation gallery and reported VSA performance use this checkpoint.

Open recommended checkpoint

Preview v1 LoRA

Adapter

A rank-64 pre-extracted LoRA that runs on top of MiniMaxAI/MiniMax-H3. Choose the vsa-datafree folder to match the recommended full checkpoint while downloading only the distilled adapter.

Open LoRA repository

VSA / Synthetic

Research ablation

Two VSA checkpoints trained on videos generated by Base H3 at steps 1300 and 1900. They share the VSA architecture and runtime requirements with the recommended checkpoint.

View FastH3 collection

Dense / Data-Free

Dense reference

The full-attention reference trained from prompts. It uses dense FA4 rather than the VSA-H3 path and exists for comparison, not as a transparent substitute for the recommended sparse checkpoint.

Open dense checkpoint

Verified quick start

How to run FastH3 locally

The official FastVideo quick start targets CUDA 13 and four NVIDIA B200 GPUs for the published profile. Other supported CUDA systems need portable flags, and the GPU count must divide H3's 56 attention heads.

1. Clone FastVideo

Create a clean Python 3.12 environment with uv, then install the FastH3 extra. The published fastvideo-kernel wheel includes the B200 sm100a VSA kernel used by the measured route.

2. Run the recommended checkpoint

Use the VSA / Data-Free model path and a structured prompt that describes both the visual sequence and the soundscape. The example below disables warmup and writes one measured generation.

3. Adapt for other CUDA systems

On other multi-GPU CUDA hardware, add --no-replicated-dit --vsa-kernel triton --no-fa4. Keep VSA enabled for the recommended checkpoint, confirm that your GPU count divides 56, and expect different latency from the B200 benchmark.

bash
git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
uv venv --python 3.12 --seed
source .venv/bin/activate
UV_TORCH_BACKEND=cu130 uv pip install \  --no-sources-package fastvideo-kernel \  -e ".[fasth3]"
bash
python examples/inference/basic/basic_fasth3.py \  --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \  --prompt "integrated_multimodal_description: A red fox runs through fresh snow at dawn. overall_soundscape: Fast pawsteps in snow, winter wind, and distant birds." \  --no-warmup \  --repeats 1

Preview boundaries

What FastH3 Preview v1 does not promise

FastH3 is an early checkpoint family. A useful evaluation starts with its published scope instead of assuming every capability of the MiniMax H3 base model carries over.

Text-to-audio-video only

Preview v1 distills the T2VA path. First-and-last-frame conditioning and omni-reference generation were not distilled in the released checkpoints. FastVideo lists FL2VA and Ref2VA as follow-up work.

Preview quality remains below Base H3

FastVideo explicitly notes remaining gaps on difficult motion, fine detail, and some audio. Four steps target minimum latency. The team is evaluating stronger training and a possible eight-step quality setting.

Published speed depends on the profile

The headline speed figures depend on FastVideo's measured combination of hardware, VSA, kernels, compilation, VAE parallelism, and warmup protocol. A different GPU or eager path should not be expected to reproduce those numbers.

License review is required

FastH3 inherits the MiniMax H3 Community License and content restrictions. Review the model license before use, redistribution, or deployment. This page is a technical guide, not legal advice.

FastH3 questions

Frequently asked questions

Short answers to the searches developers are most likely to make while FastH3 is new.

What is FastH3?

FastH3 is FastVideo's few-step checkpoint family distilled from MiniMax H3. The recommended Preview v1 model generates synchronized video and audio from text with four transformer forwards.

Is FastH3 the same as MiniMax H3?

No. MiniMax H3 is the 33B base system. FastH3 reuses the base model components but replaces or adapts the transformer with a FastVideo DMD2-distilled student designed to use far fewer denoising evaluations.

Who made FastH3?

The FastVideo team at Hao AI Lab released FastH3, with acknowledgements to Nuva Lab, NVIDIA FastGen, MiniMax, vLLM, NVIDIA, and MBZUAI. FastH3.org is an independent guide and is not the official project site.

Where can I download FastH3?

Download the recommended full checkpoint or matching LoRA from the FastVideo organization on Hugging Face. Use the official FastH3 collection to compare all released Preview v1 variants.

Is FastH3 open source?

FastVideo publishes the inference code and open weights. The model is distributed under the MiniMax H3 Community License rather than Apache 2.0, so review the license before deciding whether your intended use is allowed.

How fast is FastH3?

FastVideo reports up to 14x speedup on a single NVIDIA Blackwell GPU and less than 13 seconds for a 15-second 768p video on 8x B200 GPUs. Those are profile-specific benchmark claims, not universal latency guarantees.

Does FastH3 generate audio?

Yes. The current FastH3 preview is a text-to-video-and-audio model that generates synchronized video and stereo audio in the same pipeline.

Does FastH3 support image-to-video?

Not in the released Preview v1 distilled checkpoints. MiniMax H3 Base supports broader modes, but FastVideo states that FL2VA and Ref2VA need separate conditioning or transformer distillation and remain in development.

Which FastH3 checkpoint should I use?

Start with FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree. Use the matching vsa-datafree LoRA when you want the adapter on top of Base H3. The other checkpoints are research comparisons.

Can FastH3 run on a consumer GPU?

The weights can run on compatible hardware with enough memory, but the published fast profile uses Blackwell GPUs and specialized kernels. Portable CUDA, RTX, DGX Spark, and Apple MLX work is evolving, so check the latest FastVideo installation guide.

What does four-step mean in FastH3?

FastVideo's runner uses five sigma-grid points that produce exactly four transformer forwards. It is more precise to call this a four-forward trained schedule than to assume any arbitrary four-step scheduler configuration is equivalent.

Can I use FastH3 commercially?

Commercial suitability depends on the MiniMax H3 Community License, applicable territory, acceptable-use terms, and your deployment. Review the official license and obtain legal advice for your specific use case.

Primary evidence

Official FastH3 sources

FastH3 is changing quickly. Use these first-party pages to verify the latest checkpoint, runtime, benchmark, and license status.

FastVideo repository

Inference code, installation guides, the verified basic_fasth3.py runner, issues, and active development.

FastH3 Hugging Face collection

Recommended full weights, LoRA adapters, dense and synthetic ablations, model cards, and inherited licenses.

FastH3 Preview v1 release

FastVideo's benchmark protocol, validation gallery, checkpoint comparison, quick start, and roadmap.

Start with the recommended FastH3 checkpoint

Open the official model card, review the license and limitations, then use the verified FastVideo runner. Keep this guide nearby for checkpoint selection and portable hardware notes.

Open FastH3 on Hugging Face Read FastVideo documentation