Video Rebirth cuts MiniMax-H3 model below 14B parameters

Video Rebirth released HyperFlow H3 Flash in a GitHub repository on Nov. 6, a MiniMax-H3 derivative with 13.7 billion or 13.9 billion parameters, depending on task.

Video Rebirth said in the Nov. 6 repository that MiniMax-H3 uses about 31 billion parameters, so the release targets smaller systems for joint video-and-audio generation.

The README says the model “generates video and synchronized stereo audio together.” It lists output of 5.2 seconds to 15 seconds of video at 24 frames per second, with a default 1344×768 resolution and 32 kHz stereo audio.

Video Rebirth said HyperFlow H3 Flash supports 4, 6 or 8 sampling steps, with six as the default. A 5.2-second clip at six steps took 8.7 seconds on four B200 GPUs, 36 seconds on one H200 and 92 seconds on one RTX 5090, according to the repository.

The repository also says NVIDIA SOL-H3 runs HyperFlow H3 Flash in real time on four B-series GPUs. It describes SOL-H3 as the model’s inference-acceleration partner, meaning a system focused on speeding generation after training.

For local use, Video Rebirth lists GGUF files from 7.2 GB to 15.2 GB and an INT8 convrot option of about 16 GB. INT8 refers to 8-bit integer precision, a common compression approach that can reduce memory use.

The lowest listed footprint is 9.1 GB for Q4_K_M, and Video Rebirth says the model runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM. The repository says 16 GB GPUs have not been measured yet.

Weights are available through the VideoRebirthAI HyperFlow H3 Flash page on Hugging Face, with ComfyUI nodes and a GGUF setup guide linked from the repository.

Source: GitHub repository, Video Rebirth, Nov. 6; Hacker News post, Nov. 7