FastH3 Preview v0.1: Four Open-Weight H3 Models in Four Steps

TL;DR. FastVideo is releasing four text-to-video-and-audio FastH3 Preview v0.1 checkpoints. They reduce MiniMax H3 from 49 transformer calls to four and generate 768p video with 32 kHz stereo audio. The variants explore prompt-only and synthetic-video training, and sparse and dense attention. In our qualitative checks, fast motion and fine audio remain weaker than Base H3; first/last-frame and reference workflows are not supported yet.

This is an open release: weights, FastVideo training and inference code, exact configs, synthetic Base-H3 data, evaluation artifacts, and the sampling contract. We want the community to reproduce the recipe and improve it with us.

Why open H3 matters# #

The strongest video systems were mostly closed until MiniMax released the H3-Base weights. With downloadable weights, the community can inspect H3, post-train it, replace kernels, and run it on its own hardware.

That ecosystem is already growing. On August 26, fal announced H3 Max, a hosted post-trained H3 model. fal reports stronger quality and a latency below three seconds for a five-second 768p request. We have not independently measured those claims, and fal has not announced downloadable H3 Max weights or training code. FastH3 takes a complementary path by making the acceleration recipe a shared FastVideo asset, not only an endpoint.

This continues our work on FastWan sparse distillation and FastWan-QAD. Both releases paired checkpoints with their FastVideo training and inference stacks. FastH3 brings the same open development model to H3.

FastVideo code uses Apache 2.0. H3 and FastH3 weights use the custom MiniMax H3 Community License, not a standard open-source model license. Please read it before using or redistributing the weights.

Four preview checkpoints# #

FastH3 distills H3’s base transformer for text-to-video-and-audio (T2VA) and reuses the H3-Base text encoder, video VAE, audio VAE, tokenizers, and schedulers. All four checkpoints belong to Preview v0.1; their labels describe the recipe, not a version order.

Alongside the weights, FastVideo is releasing:

• DMD2, backward-simulation, VSA, and dense training code with exact configs.
• Prompts and synthetic Base-H3 videos, with manifests and provenance.
• Inference scripts, evaluation prompts, timing logs, votes, and negative ablations.

Another team should be able to reproduce the recipe and train the next checkpoint without private code or data.

VSA / Synthetic / Step 1300VSA / Synthetic / Step 1900Dense / Data-Free“Data-free” means training uses prompts but no target videos. The synthetic variants use videos generated by Base H3, not camera footage, and come from one run at two steps. The dense variant provides a full-attention comparison. These runs use different training lineages, so their quality differences are useful signals, not controlled ablations. The 8-step and low-noise runs are separate experiments.

All four checkpoints support T2VA only. H3 uses the base transformer for both T2VA and first/last-frame-to-video-and-audio (FL2VA), but these students were not trained with first/last-frame conditioning. Reference-to-video-and-audio (Ref2VA) uses a separate reference transformer and needs its own distilled checkpoint.

Performance# #

We will fill this table with results from the exact public checkpoints. Each local test uses FastVideo’s default optimized path and H3’s nominal five-second shape: 124 frames at 24 fps (about 5.17 seconds), 1344x768, stereo audio, and guidance 1.0. We report the median of three timed requests after one full warmup. Model and compilation are excluded. End-to-end time includes encoding, denoising, decoding, audio, muxing, and file output; denoise time measures only the DiT loop.

The three VSA checkpoints share one latency row because their tensor shapes, four-step schedule, sparse layout, and FastVideo inference graph are identical. On B200, the default path uses FastVideo’s tile-64 CUDA VSA kernel, regional DiT compilation, H3 fusions, and compiled parallel video VAE. We will evaluate the three checkpoints’ quality separately. H3 Max is a hosted service on undisclosed hardware, so fal’s number is not an apples-to-apples local result.

How FastH3 works# #

Base H3 calls its 33B audio-video diffusion transformer 49 times. FastH3 lowers that cost in two ways: four calls instead of 49, and less attention work inside each call.

DMD2 cuts the number of calls. DMD2 trains a student against a frozen Base H3 teacher and a learned critic. The student learns from the gap between their scores. For prompt-only runs, backward simulation exposes the student to the few-step states it will see at inference. The synthetic-video runs instead start from forward-noised Base-H3 video-and-audio latents.

VSA makes each call cheaper. The VSA paper introduces trainable sparse attention for video diffusion. In FastH3, the student keeps about 10% of eligible video-to-video tiles (90% sparsity) with 64-token blocks. Text and audio remain dense. The teacher and critic also remain dense, giving the sparse student a full-attention target. This extends the FastVideo sparse-distillation recipe to H3.

Reference conditioning needs a new policy. Keeping a full reference video dense would erase most of VSA’s benefit. Our reference-aware policy treats each reference video and the output video as a separate sparse region, while text, images, and audio stay dense. This work uses H3’s separate reference transformer and is not part of Preview v0.1.

Where four steps still fall short# #

Fast motion and the noise schedule#

In our qualitative checks, FastH3 is strongest on slower scenes and simple motion. In faster scenes, hair, water, grass, small objects, and sharp moving edges can become soft or smeared.

The four-step ladder looks evenly spaced: [999, 749, 500, 250]

. But H3 maps each base timestep through a shift of 12 for video and 3 for audio. The effective noise levels are:

The current model’s last call happens while video noise is still 0.80. One large final update must recover fine detail and fast motion at once. That matches the blur and smearing we see, but it is still a hypothesis.

The 8-step run adds more calls. The low-noise run keeps four calls but moves the last one closer to the clean output, from video noise 0.80 to 0.387. FastVideo uses the same schedule for training and validation because changing only the inference schedule can hurt a distilled model. More low-noise work may sharpen edges but weaken composition or large motion, so we will compare both runs before choosing a new default.

Aligning audio and video#

We fixed two sources of mismatch. First, FastVideo now derives every video and audio noise level from one shared base timestep, then applies H3’s video shift of 12 and audio shift of 3. Training, validation, and inference stay on the same paired schedule.

Second, H3 packs audio on a 40 Hz clock tied to the video length, while the audio VAE can round up by one latent. FastVideo now trims or edge-pads the audio latent to the exact packed length. These are latent and diffusion-clock fixes, not a late MP4 offset. Fine audio detail can still be weak, so evaluation will include listening tests and an audio-video sync metric.

Try FastH3# #

The optimized FastVideo path targets NVIDIA B200 GPUs with CUDA 13. Startup loads the model, and the first generation compiles it. Use a warm request for steady-state timing.

git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
git checkout FASTVIDEO_RELEASE_COMMIT
uv venv --python 3.12 --seed
source .venv/bin/activate
UV_TORCH_BACKEND=cu130 uv pip install -e ".[fasth3]"
PROMPT='integrated_multimodal_description: A red fox runs through fresh snow at dawn. overall_soundscape: Fast pawsteps in snow, winter wind, and distant birds.'
MODEL_PATH='FastVideo/FastVideo-FastH3-4-step-v1.2'
export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250

python examples/inference/basic/basic_fasth3.py \
  --model-path "$MODEL_PATH" \
  --prompt "$PROMPT" \
  --output outputs/fasth3_fox \
  --num-gpus 4 \
  --repeats 1

The example uses five scheduler points, which produce exactly four DiT calls. Keep the trained schedule, 90% VSA sparsity, tile size 64, and guidance 1.0; changing them can hurt quality. This command uses VSA / Synthetic / Step 1900. The other VSA checkpoints use the same command with a different model path. See the Dense / Data-Free model card for its command.

What comes next# #

Preview v0.1 is an early checkpoint family, not a claim that four steps have solved H3. FastVideo’s next priorities are:

1. Publish a quality result people can audit#

We will compare every checkpoint with Base H3 and H3 Max on 60 held-out prompts. The study will match the 768p output tier, aspect ratio, and audio settings as closely as each system allows; record the actual frame count and duration; randomize left and right; and collect at least three votes per pair. We will publish checkpoint hashes, prompts, seeds, every vote, prompt-level confidence intervals, and representative wins, ties, and losses with audible clips.

For the direct H3 Max comparison, we will set its prompt_expansion_mode to

disabled

, use its seed control, and time the full client request. We will report the hosted default with prompt expansion as a separate product comparison. FastH3 clips will use FastVideo’s default optimized path. Preference is (wins + ties / 2) / votes

; confidence intervals will be computed over prompts rather than treating repeated votes as independent samples.| FastH3 candidate | Opponent | FastH3 W / T / L | Preference (95% CI) | Prompts | Votes / pair | |---|---|---|---|---|---| | VSA / Data-Free | Base H3 | TBD | TBD | 60 | TBD | | VSA / Data-Free | H3 Max | TBD | TBD | 60 | TBD | | VSA / Synthetic / Step 1300 | Base H3 | TBD | TBD | 60 | TBD | | VSA / Synthetic / Step 1300 | H3 Max | TBD | TBD | 60 | TBD | | VSA / Synthetic / Step 1900 | Base H3 | TBD | TBD | 60 | TBD | | VSA / Synthetic / Step 1900 | H3 Max | TBD | TBD | 60 | TBD | | Dense / Data-Free | Base H3 | TBD | TBD | 60 | TBD | | Dense / Data-Free | H3 Max | TBD | TBD | 60 | TBD |

We will publish losses and ties, not only wins. Reviewers will see the full clip and hear its audio. Four padding duplicates from distributed validation are excluded.

2. Improve motion and offer an eight-step option#

We will finish the four-step low-noise A/B and the 8-step control. We will then test stronger final-step training, more low-noise critic samples, motion-sensitive losses, and learned timestep placement. The four-step model targets minimum latency; eight steps may be a better quality setting. We will release an 8-step checkpoint only after a matched comparison.

3. Add FL2VA and Ref2VA#

T2VA is only one H3 workflow. FL2VA uses the base transformer but needs new conditioning training. Ref2VA uses the separate reference transformer, so FastVideo must distill it separately. We will evaluate mixed reference types, long clips, and reference fidelity before releasing either workflow.

4. Evaluate audio quality#

A decodable 32 kHz stereo track is only the first gate. We will measure speech, effects, ambience, music, timing, and audio-video sync, and add human listening tests. The paired-schedule and packed-length checks will remain regression tests.

5. Make H3 easier to run and extend#

At four calls, encoding, VAE decode, audio, and file output become a larger share of latency. FastVideo will keep improving VAE compilation and parallelism, sparse compilation, portable kernels, cold start, and multi-GPU serving. We will also study FP8 and NVFP4 with visible quality gates. Stable releases will include the checkpoint, config, sampling contract, code revision, data provenance, and negative ablations.

Help us test more hardware# #

Our published numbers use NVIDIA B200 GPUs. We welcome results from other single- and multi-GPU systems that can load FastH3. Run the default example and report the GPU model and count, driver and CUDA versions, FastVideo commit, checkpoint, output shape, warmup, end-to-end and denoising time, peak memory, and a sample clip.

Share results, unsupported hardware, regressions, and new ideas on GitHub or in the FastVideo Slack. Explore the four-checkpoint collection, run the FastVideo example, and show us where FastH3 works—and where it does not.

Acknowledgements# #

We thank our collaborators on the NVIDIA FastGen team. Their reference implementation helped us align DMD2’s score clock, modality shifts, and backward-simulation behavior. We also thank MiniMax for releasing H3-Base, fal for adding another H3 direction to the ecosystem, the DMD2 and VSA authors, the FlashAttention and CUTLASS teams, and every FastVideo contributor who built and tested the model, kernels, training path, evaluation tools, and serving stack.

Citations# #

Method background: the DMD2 paper, the VSA paper, our earlier FastWan sparse-distillation blog, and the FastWan-QAD blog. If you build on FastH3, please cite this release, DMD2, VSA, and FastVideo.

@misc{fastvideo_fasth3_2026,
  title        = {FastH3 Preview v0.1: Four Open-Weight H3 Models in Four Steps},
  author       = {FastVideo Team},
  year         = {2026},
  howpublished = {\url{https://haoailab.com/blogs/fasth3-preview/}},
}

@misc{yin2024improved,
  title         = {Improved Distribution Matching Distillation for Fast Image Synthesis},
  author        = {Tianwei Yin and Michaël Gharbi and Taesung Park and Richard Zhang and Eli Shechtman and Fredo Durand and William T. Freeman},
  year          = {2024},
  eprint        = {2405.14867},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
}

@article{zhang2025vsa,
  title   = {VSA: Faster Video Diffusion with Trainable Sparse Attention},
  author  = {Peiyuan Zhang and Yongqi Chen and Haofeng Huang and Will Lin and Zhengzhong Liu and Ion Stoica and Eric Xing and Hao Zhang},
  journal = {arXiv preprint arXiv:2505.13389},
  year    = {2025},
}

@software{fastvideo2024,
  title  = {FastVideo: A Unified Framework for Accelerated Video Generation},
  author = {The FastVideo Team},
  url    = {https://github.com/hao-ai-lab/FastVideo},
  year   = {2024},
}