Video Rebirth cuts MiniMax H3 to 14B in six steps

Video Rebirth released HyperFlow H3 Flash on November 6, 2026, a compressed version of MiniMax's H3 video-and-audio model that generates video in six sampling steps at 13.7 billion parameters.

The release is open weights: code on GitHub under Apache 2.0, weights on Hugging Face, and a showcase on Video Rebirth's site. The maker calls it the third release in its HyperFlow series.

The compression is structural. HyperFlow H3 Flash drops three of MiniMax H3's 50 transformer blocks and factors 102 linear layers to reach 13.7 billion parameters, down from the base model's roughly 31 billion. It quantizes 180 layers to INT8 behind a Hadamard rotation, then adds two-time (t, r) conditioning so the sampler can converge in four, six or eight passes without classifier-free guidance. Six is the default. It still generates video with synchronized stereo audio.

Output runs 5.2 to 15 seconds at 24 frames per second, 1344x768 by default, with a 32 kHz stereo soundtrack. Video Rebirth's benchmarks list 8.7 seconds for a 5.2-second clip on four B200 GPUs at six steps, 36 seconds on one H200 and 92 seconds on an RTX 5090. The smallest GGUF build, Q4_K_M, is 9.1 GB, and the maker says the checkpoint runs on 24 GB and 32 GB consumer GPUs in PCs with 16 GB of system memory. Sixteen-gigabyte GPUs are not measured yet.

For real time, Video Rebirth points to NVIDIA's SOL-H3, which it names as its inference-acceleration partner. SOL-H3 runs HyperFlow H3 Flash on four B-series GPUs. Video Rebirth says a single B200 gets a 12.7x speedup over the H3 baseline.

Video Rebirth is a Singapore startup founded by Wei Liu, who left Tencent in late 2024 after helping lead its Hunyuan foundation-model effort, as RuntimeWire notes. HyperFlow H3 Flash comes from a different team than the other H3 compressions already out, including a four-step distillation by separate makers.

The catch is the license. HyperFlow H3 Flash is a model derivative of MiniMax H3 and ships under the MiniMax H3 Community License Agreement. The model card states the agreement licenses the weights worldwide "excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America," and that people and organizations in those places must get MiniMax's prior authorization. Products over $20 million in annual revenue need written authorization too, and publicly distributed output must be disclosed as AI-generated.

For creators, that split is the story. The engineering is downloadable today, and a 9.1 GB checkpoint fits a card most people already own. Whether you can legally build on it depends on where you are.

HyperFlow H3 Flash is on GitHub and Hugging Face; the ComfyUI nodes and the GGUF setup guide sit in the same repository.