FastH3 Preview v1: 4-Step MiniMax H3 Distillation From FastVideo

FastVideo, in collaboration with Nuva Lab and the NVIDIA FastGen team, open-sourced FastH3 Preview v1, a 4-step distilled version of MiniMax H3 for text-to-video-and-audio generation (checkpoint | blog | GitHub). It replaces the base model's 49 transformer calls with 4 and adds 90% sparse attention, reaching up to 14x speedup on a single NVIDIA Blackwell GPU.

FastH3 Preview v1: open-weight 4-step sparse distilled MiniMax H3 from FastVideo

Benchmarked speed

On B200 hardware at 1344x768 and 24 FPS with audio, the recommended VSA / Data-Free checkpoint generates a 15-second clip in 47.2 seconds on one B200 (14.38x faster than the dense base model) and 15.5 seconds on eight B200s, with 5s and 10s clips at 16.2s and 31.1s on a single GPU. Timings cover the full pipeline: encoding, denoising, decoding, audio, muxing and file output.

How the distillation works

FastH3 cuts cost in two ways:

• Four calls instead of 49. Distribution Matching Distillation (DMD2) trains a student against a frozen Base H3 teacher with a learned critic. Prompt-only runs use backward simulation; a synthetic-data variant starts from forward-noised Base-H3 video-and-audio latents.
• Less attention work per call. The student adopts VSA (trainable sparse attention for video diffusion) at 90% sparsity, running on FastVideo's tile-64 CUDA VSA kernel with regional DiT compilation, H3 fusions and a compiled video VAE.

What is in the release

The checkpoints are trained and validated at multiple aspect ratios, including square, portrait, landscape and ultrawide 768p, with custom heights and widths in multiples of 32. They reuse the H3-Base text encoder, video VAE, audio VAE, tokenizers and schedulers.

Scope and ComfyUI availability

Preview v1 covers text-to-video-and-audio only: FL2VA (first-frame image conditioning) and Ref2VA (reference-image conditioning) were not distilled, and difficult motion, fine detail and some audio can fall below the base model. FastVideo lists image-reference support, NVFP4 quantization, RTX / DGX Spark / Apple MLX optimizations, and new runs with NVIDIA's PDD method on the roadmap.

For ComfyUI users today, FastH3 is a FastVideo-stack release rather than a drop-in checkpoint: the VSA adapters need FastVideo's VSA-H3 backend and launchers, not a generic LoRA loader. ComfyUI-native acceleration for H3 currently comes from distillation LoRAs such as H3 Turbo-SLA and PDD Acc LoRA, which run on standard ComfyUI H3 workflows. A ComfyUI-compatible FastH3 path is worth watching as the team brings the models to consumer hardware.