FastH3 Trim: MiniMax H3 on an 8 GB GPU, and who may use it

FastH3 Trim is an experimental, pruned version of MiniMax H3 that keeps 42 of the original 50 transformer blocks and samples in 8 steps, so video with synchronized audio can run on a single consumer NVIDIA GPU; one write-up reports it working on a memory-capped 8 GB RTX 4090. It is released under the MiniMax H3 Community License, which has territory and revenue conditions you should read before you ship clips made with it.

What was cut from MiniMax H3?

According to the model card, the FastVideo team removed eight blocks whose removal changed the video and audio predictions least, replaced each block's timestep conditioning with a shared rank-16 basis, and distilled the result to eight sampling steps. The card calls it 4.2 times smaller than base H3 and faster than FastH3 V2.

The same page labels it an experiment. The ComfyUI Wiki write-up quotes the developers saying that removing blocks makes the model faster and smaller but costs some quality. Treat it as a draft engine, not a replacement for the full model.

Which file fits which GPU?

The ComfyUI repack ships four formats. It needs ComfyUI 0.36.0 or later, the FastVideo FastH3 text-to-video template, a Qwen3VL text encoder and separate audio and video VAE files.

The 8 GB figure is not a file size. The NVFP4 file is 11.9 GB, so the 8 GB case relies on the memory-capped test the ComfyUI Wiki describes, where performance degrades. The same write-up gives roughly 43 to 83 seconds for a 5-second 480p clip on an RTX 4090, depending on the memory cap.

Can you use the clips commercially?

The repack lists the minimax-h3-community-license-agreement. In the license text I read, use is limited to an Applicable Territory that excludes the EU, UK, South Korea and the USA, and it says you may not use the works or their Outputs outside that territory. Companies above $20 million in annual revenue need separate written authorization from MiniMax. Users must prominently display "MiniMax H3" on product interfaces, and Outputs cannot be used to improve another AI model.

That is a summary, not legal advice. If your team or your viewers are in an excluded region, get the license reviewed or contact MiniMax before you build on the weights.

When is a hosted workflow the practical choice?

If the box under your desk is not an NVIDIA card, or you need a clip you can ship without a territory check on your own infrastructure, a hosted route is simpler. Sume's Video Router lists minimax-h3, which the docs say accepts 5 to 15 seconds at native 480p or 768p. Sume bills at the provider list price with a 1.25 house margin, and you read the live capabilities from the catalog rather than assuming them.

That is a different product from FastH3 Trim: it is the hosted H3 row, not these pruned weights, and it has no local setup. Run Trim when you want free local drafts on your own GPU and have cleared the license. Call the hosted row when you want the clip and nothing else to maintain.