Video Rebirth publishes HyperFlow H3 Flash, a compressed MiniMax-H3
Original title: VideoRebirthAI/hyperflow-h3-flash — new model trending #30 on Hugging Face
The 13.7B–13.9B model makes 5.2–15 second video with stereo audio in 4–8 steps, under a restricted MiniMax license.
Video Rebirth released HyperFlow H3 Flash on Hugging Face, where it was trending at number 30. It is a compressed, few-step derivative of MiniMax-H3 that generates video and synchronized stereo audio, at 13.7B or 13.9B parameters versus about 31B for the base transformer. Default output is 1344×768 at 24 fps; Q4_K_M weights are 9.1 GB. The card says the MiniMax H3 Community License is not granted in the EU, UK, South Korea, or the US without prior MiniMax authorization, and that the repo is not an official MiniMax product.
Unconfirmed: Speed and size figures are the publisher's own measurements, not independently verified.
- Tasks listed are t2va, fl2va, and ref2va, the last taking up to 12 references.
- Default sampling is 6 steps without classifier-free guidance.
- Card cites 8.7 seconds for a 5.2-second clip on 4× B200 at 6 steps.
- NVIDIA SOL-H3 is named as a day-0 inference partner for 4× B-series GPUs.
- Companion text encoder and VAEs load from upstream MiniMax-H3 repositories.
- Vendors
- Video Rebirth
- AI models
- HyperFlow H3 FlashMiniMax-H3
From the original article
12 min read at huggingface.coimage-text-to-video · 64,221 downloads · 138 likes · license other
By VideoRebirthAI · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.