FastH3: Free Fast MiniMax H3 Video with Sound

FastH3 is FastVideo’s free 8-step version of MiniMax H3 that makes video with sound from photos, clips and voice. See speed, hardware and licence limits.

The FastVideo team at Hao AI Lab (UC San Diego) released FastH3 Omni on October 20, 2026, and a local version for desktop GPUs and Macs a day later. It’s a distilled copy of MiniMax H3 that needs only 8 denoising steps, which is why it’s so much faster than the original. You give it a prompt plus reference images, video clips or a voice sample, and it returns a finished video with synchronized sound. We checked the release post, the local AI post and the model card on October 23, 2026.

Is FastH3 free?

Yes, the weights are free to download from Hugging Face. But it isn’t a normal open-source licence. FastH3 inherits the MiniMax H3 Community License, which has real limits on where and how you can use it (see Limitations below). There’s no paid app or subscription, so your cost is the GPU you run it on.

FastH3 Pricing (2026)

FastH3 costs $0 to download. What you pay for is hardware, and the speed difference between machines is huge. Here’s how long the FastVideo team’s own test took, a 5-second 480p clip from one reference photo:

Hardware	Time for a 5-second clip
8x NVIDIA GB200 (768p)	5.7 seconds
NVIDIA RTX 5090	62 seconds
NVIDIA DGX Spark	192 seconds
NVIDIA RTX 3090	16.9 minutes
Apple M5 Max (64 GB)	19.5 minutes

So an RTX 5090 makes it practical at home, while older 24 GB cards and Macs work but are slow. If you just want video with sound without a GPU, a paid app like Kling 4.0 or Seedance Pro is easier.

What FastH3 can do

• Reference-to-video with audio: 24 fps video with 32 kHz stereo sound generated together.
• Lots of references: text plus up to 9 images, 3 video clips and 3 audio clips, up to 12 in total. You need at least one image or video.
• Fast: 5.6 to 11.9 times lower latency than the base MiniMax H3 on 8 GB200 GPUs. One example went from 35.44 seconds to 5.67 seconds.
• Text to video too: 52.6 seconds on a DGX Spark with no reference images.

How to run FastH3 locally

FastVideo publishes builds for different hardware:

1. NVIDIA Blackwell (RTX 50 series, DGX Spark): the NVFP4 QAD checkpoint is the fastest option.
2. RTX 4090 and other 24 GB cards: use the FP8 build.
3. Mac: an MLX INT6 build runs on Apple Silicon.
4. ComfyUI: BF16, INT8, FP8 and NVFP4 checkpoints for node workflows.

The model is big: a transformer of about 35B parameters plus a Qwen3-VL text encoder of about 32B, so plan for plenty of disk space and memory. For other open models that fit consumer GPUs, see the best local AI models, and for free online options see our list of the best free AI text to video generators.

Limitations

• Region limits: the MiniMax H3 Community License excludes the EU, UK, South Korea and USA.
• Revenue cap: companies earning over $20 million a year need written permission from MiniMax.
• Attribution: commercial products must show “MiniMax H3” prominently in their interface.
• No training on outputs: you can’t use its videos to improve other models.
• Slow on older hardware: about 17 to 20 minutes per clip on an RTX 3090 or M5 Max.

Verdict

FastH3 is the fastest way to run MiniMax H3-quality video with sound on your own GPU, and on an RTX 5090 it’s genuinely usable. Read the licence first: if you’re in an excluded region or building a large commercial product, it’s not for you.

FastH3 makes MiniMax H3 video with sound fast enough to use on a single RTX 5090.

It’s free to download, but check the MiniMax H3 licence’s region and revenue limits before you build anything on it.