MiniMax drops FastH3 v1 for ultra-fast open video

FastH3 v1 is an open-weight text-to-video-and-audio model designed for extreme speed. Running on eight NVIDIA B200 GPUs, it churns out a 15-second 768p video in less than 13 seconds, clocking up to a 14x speedup on a single Blackwell chip. If you want production-grade video generation without waiting minutes per clip, this is built for you.

Developed jointly by Hao AI Lab and MiniMax alongside NuvaLab and NVIDIA FastGen, FastH3 builds on MiniMax's H3-Base architecture. The team spent over 1,000 B200 GPU training hours post-training the model. Best of all, you do not need separate checkpoints for different formats: a single file handles variable resolutions, aspect ratios, and durations right out of the box.

How it hits sub-realtime speeds

The secret sauce lies in post-training distillation and specialized attention kernels. The primary FastH3 Preview v1 checkpoint relies on a 4-step data-free approach, training strictly from text prompts without requiring ground-truth target videos.

To slash latency, it implements FastVideo's Video Sparse Attention (VSA-H3) backend, running at 90% sparsity with a tile size of 64. Instead of computing heavy dense attention across every token, it skips the redundant work while preserving visual quality and audio sync. FastH3 reuses the proven H3-Base text encoder, tokenizers, and video/audio VAEs, distilling only the base transformer.

It is also not locked into standard 1344x768 widescreen. FastVideo accepts any custom dimensions in multiples of 32, comfortably handling square, vertical, landscape, and ultrawide formats.

What is available and what is missing

The team has released full model weights, a pre-extracted LoRA, synthetic-data ablations, and working inference code. The complete training recipes will drop soon.

There are a few trade-offs to keep in mind. Preview v1 currently handles text-to-video-and-audio only. Conditioning on first and last frames or image references is still in the works and slated for release in the coming weeks. Furthermore, hitting that 14x benchmark requires NVIDIA Blackwell silicon and the custom VSA kernel—dense attention cannot simply serve as a drop-in substitute.

The developers plan to introduce NVFP4 quantization and bring optimizations to consumer RTX cards, DGX Sparks, and Apple MLX soon. If you have the hardware, you can grab the weights and inference code directly from FastVideo's repository.

Why it matters

We think FastH3 is a big deal because it takes open-source video from an academic curiosity to something that can survive commercial infrastructure budgets. Generating video faster than realtime removes the biggest bottleneck standing between generative video and live production pipelines.

If you run video infrastructure or want to integrate interactive generation into your product, you should test this checkpoint now. Just keep in mind that hardware requirements remain steep until lightweight quantizations land on consumer hardware.