Video Rebirth shrinks MiniMax-H3 to 13.7B so it runs faster on consumer PCs

How it works

A compact AI model that turns text into short video clips with synced sound, running on home gaming graphics cards.

What you could do with it

A hobbyist could generate a short video clip with synced stereo audio on their own consumer-grade graphics card like an RTX 5090 without paying for cloud computing ¹.
A developer could create 5.2 to 15-second videos directly from text descriptions or reference images and play them back with matching sound in ComfyUI or via a Python package ¹.
A team running data-centre hardware could use up to four NVIDIA Blackwell B-series GPUs together to render clips in real time through NVIDIA's SOL-H3 acceleration framework ¹.

An exampleInvented to explain it

Maya, editing a YouTube vlog on her weekend PC

MayaMake a clip of waves rolling on a sunset beach

HyperFlow H3 FlashGenerating 5.2 seconds at 24 fps with stereo sound...

MayaHow long will that take on my RTX 5090?

HyperFlow H3 FlashAbout 92 seconds with six refinement steps.

MayaGot it, let's see the result.

HyperFlow H3 FlashDone. Here is your beach video with wave sounds.

Maya watched the clip on her own desktop without paying cloud bills.

What the model does

HyperFlow H3 Flash takes a text prompt, one or more reference images, or both and produces 5.2 to 15 seconds of video at 24 frames per second (the default resolution is 1344 by 768 pixels) alongside a 32 kilohertz stereo soundtrack ¹. Unlike some generators that produce silent footage and add music later, the video and audio are created together in a single pass, so the sounds match what happens on screen .

The model works with four, six or eight "steps" — the number of times the AI refines its output — and skips classifier-free guidance (CFG) ¹.

How small it is and how fast

MiniMax-H3 originally had roughly 31 billion parameters; the Flash variant has just 13.7 billion (or 13.9 billion for the version that handles multiple reference inputs) ¹. That cuts the base transformer weight file from 62 gigabytes to about 16 gigabytes in INT8 format .

Because the weights are available in several compressed formats, users can pick the version that fits their GPU ¹:

GGUF files range from 7.2 GB (Q3_K_M) to 15.2 GB (Q8_0) and run on ComfyUI, Unsloth Studio or stable-diffusion.cpp, including on 16 GB and 32 GB graphics cards ¹.
INT8 safetensors sit at around 16 GB and are meant for 24 GB+ RTX GPUs that have at least 16 GB of system RAM, as well as for data-centre GPUs when paired with NVIDIA's SOL-H3 acceleration library ¹.

NVIDIA's SOL-H3 framework can already run HyperFlow H3 Flash in real time on four Blackwell B-series GPUs ¹. On a single H200, generating a 5.2-second clip takes about 36 seconds at the default six steps; on four B200 cards the same clip finishes in under nine seconds .

Guess

How much does the full Home setup download?

How to get it

All weights live in the VideoRebirthAI/hyperflow-h3-flash folder on Hugging Face, and a full setup guide is in the companion GitHub repository ¹. Users can install the Python package directly from GitHub (pip install git+https://github.com/Video-Rebirth/hyperflow-h3-flash) or run pre-built templates inside ComfyUI . First-time downloads total roughly 90 GB (or 68 GB with FP8 text-encoder precision), covering the companion text encoder, VAEs and the INT8 transformer; the full 62 GB MiniMax-H3 base model is never downloaded .

Because the release includes ready-made ComfyUI workflows and command-line tools, a person with a 24 GB GPU and 16 GB of system RAM can start producing short video clips at home without writing custom code ¹.

How good is it

On a test generating a 5.2-second clip at six steps, a single H200 GPU produced the result in about 36 seconds, while four B200 cards finished the same task in under nine seconds according to the project page ¹. A single RTX 5090 took roughly 92 seconds for that same clip, meaning consumer hardware is slower but still functional . NVIDIA's SOL-H3 framework can already push HyperFlow H3 Flash into real-time playback when four Blackwell B-series GPUs are used together, according to the announcement . No external rival benchmarks were provided.

Key facts

Model: HyperFlow H3 Flash (13.7 B params, compressed from ~31 B MiniMax-H3)
Output: 5.2–15 s video at 24 fps (1344×768) with synced 32 kHz stereo audio
Speed: 8.7 seconds using four B200s; 36 seconds with one H200; 92 seconds on one RTX 5090
Formats: GGUF (7.2–15.2 GB), INT8 (~16 GB); runs on 24/32 GB consumer GPUs
Source: Video Rebirth GitHub repository and Hugging Face hub