HyperFlow H3 Flash puts real-time MiniMax video on consumer GPUs

HyperFlow H3 Flash is an open-source audiovisual model from Video Rebirth that shrinks MiniMax-H3 from roughly 31B parameters down to 14B. Best of all, it generates video alongside synchronized stereo sound in one pass.

The real headline here is speed and memory efficiency. By cutting out classifier-free guidance entirely, the model needs just 4 to 8 sampling steps (defaulting to 6). It outputs 5.2 to 15 seconds of 24 fps video at 1344×768 resolution, complete with a 32 kHz stereo audio track. Powered by NVIDIA SOL-H3 on four Blackwell B200 GPUs, it actually hits real-time generation matched to playback speed. On a single RTX 5090, a 5.2-second clip takes just 92 seconds.

Slimmed down for desktop GPUs

Original MiniMax-H3 setups demand enterprise hardware, largely because the base transformer alone sits at over 60GB. Video Rebirth got it down to size by cutting structural layers, applying INT8 quantization, and switching to few-step convergence.

The resulting checkpoint lands at 13.7B parameters for first-and-last frame tasks (fl2va) and 13.9B for reference-guided tasks (ref2va). Shipped in GGUF format, the Q4_K_M quant drops down to just 9.1 GB, and the smallest Q3_K_M build sits at 7.2 GB. That means anyone with a 24GB VRAM card like an RTX 3090 or RTX 4090 and 16GB of system RAM can run it locally.

It handles three main modes: text-to-video-and-audio, keyframe-to-video, and complex reference conditioning. The reference pipeline can take up to 12 items, mixing up to 9 images, 3 video clips with their existing audio tracks, and 3 audio snippets directly aligned to prompt tokens.

How to get it running

Code is up on GitHub under Video-Rebirth/hyperflow-h3-flash, while weights are hosted on Hugging Face. Getting started depends mostly on your hardware.

For consumer PCs with 64GB of system RAM or less, use GGUF. You can plug it directly into ComfyUI with the ComfyUI-HyperFlow-Flash custom node, run it through Unsloth Studio, or use stable-diffusion.cpp. If you are running on a 16GB or 32GB system, make sure to add the --fast-disk flag to avoid memory crashes on longer clips.

If you have a workstation with 69GB or more system RAM, or access to data-center chips, the Python CLI and INT8 convrot format run 10% to 20% faster than GGUF. An RTX 4090 completes a 5.2-second clip in 130 seconds using INT8 offloading.

Why it matters

We think HyperFlow H3 Flash solves one of the biggest bottlenecks in open AI video: bringing native, synchronized sound into reach without an entire server rack.

Most video models force creators to generate silent frames first and cobble together background audio in post-production. MiniMax proved unified generation works, but its raw compute demands were out of reach for individual creators. Squeezing a functional 768p audiovisual pipeline into a 9.1 GB file is a major practical win for animators and indie developers looking to build local workflows.

Keep two caveats in mind before diving in. First, the weights operate under the MiniMax H3 Community License Agreement, which restricts commercial use in jurisdictions like the US, EU, UK, and South Korea without prior clearance. Second, pushing out a full 15-second generation will still push a 24GB consumer card right to its thermal and memory limits.