Deploying compressed 13.7B video and audio models on consumer GPUs

Video Rebirth has released HyperFlow H3 Flash, a compressed version of the MiniMax-H3 model designed for joint video and audio generation. Published on November 7, 2026, this release enables developers to run 13.7 billion parameter models on consumer-grade GPUs by employing aggressive structural pruning and INT8 quantization. The project provides both a Python package for direct integration and GGUF support for use in tools like ComfyUI.

What happened

The release introduces two primary model variants: one for text-to-video-and-audio (t2va) and another for reference-based generation (ref2va). These models are derived from the original MiniMax-H3 architecture but have been significantly reduced in size and computational demand. The t2va variant contains 13.7 billion parameters, while the ref2va variant holds 13.9 billion. Both are available in INT8 format for high-performance servers and GGUF format for broader compatibility with consumer hardware.

The engineering team behind Video Rebirth focused on making these large models accessible without requiring data-center infrastructure. They achieved this by dropping specific transformer blocks and applying low-rank factorization to linear layers. The release includes detailed benchmarks showing performance across various NVIDIA GPUs, from the RTX 3090 to the latest H200 and B200 data-center cards. It also provides a command-line interface and a Python API for seamless integration into existing pipelines.

How it works

HyperFlow H3 Flash reduces the computational load through three main techniques. First, it uses structural pruning, removing three out of fifty transformer blocks in the t2va model and one block in the ref2va model. Second, it applies low-rank factorization to over one hundred linear layers, further reducing parameter count. Third, it employs rotated W8A8 INT8 quantization, which compresses weights and activations while maintaining precision through a block-diagonal Hadamard rotation.

To manage memory constraints on smaller GPUs, the system supports multiple offload modes. In "stream" mode, only the active VAE and one transformer block or text-encoder layer remain in VRAM at any time, allowing the model to run on cards with as little as 24 GB of memory. This mode is only about 2% slower than keeping all components in VRAM. Additionally, the model uses a two-time conditioning mechanism, where every token receives information about both the current noise level and the step endpoint, enabling high-quality generation in just four to eight sampling steps.

Key details

• Model Size: The t2va model has 13.7 billion parameters, and the ref2va model has 13.9 billion parameters after pruning and factorization.
• Sampling Steps: Users can choose between 4 steps for speed, 6 for balance, or 8 for quality, with sigma grids provided in the schedule module.
• Memory Optimization: The "stream" offload mode allows running 15-second mixed-reference clips on 24 GB VRAM GPUs by bounding activation memory with block chunking.
• Text Encoder Efficiency: Using FP8 precision for the Qwen3-VL text encoder saves approximately 22 GiB of system RAM when offloading, requiring the fp8 extra installation.
• Multi-GPU Support: The pipeline supports Ulysses context parallelism, splitting transformer sequences across up to eight GPUs for faster inference and lower per-GPU memory usage.
• License Restrictions: The model weights are distributed under the MiniMax H3 Community License Agreement, which restricts use in the EU, UK, Republic of Korea, and US without authorization.

Why it matters

For software engineers and technical founders, this release lowers the barrier to entry for high-fidelity video and audio generation. Previously, running models of this scale required expensive data-center GPUs with hundreds of gigabytes of VRAM. By optimizing for consumer hardware, Video Rebirth enables local development and deployment on workstations equipped with RTX 4090 or even RTX 3090 cards. This shift allows for more iterative experimentation and reduces reliance on cloud inference services, which can be costly and introduce latency.

The availability of GGUF formats also integrates these models into the broader open-source ecosystem, particularly within ComfyUI. This compatibility means developers can leverage existing workflows and community nodes rather than building custom inference engines from scratch. The detailed benchmarks provided help teams make informed decisions about hardware allocation, balancing generation speed against memory constraints based on their specific production requirements.

What you can do

• Install the Python package using make install to set up the virtual environment and dependencies for INT8 runtime execution.
• Use the --offload stream flag when running on GPUs with less than 60 GB of VRAM to manage memory usage effectively without significant speed loss.
• Enable FP8 precision for the text encoder by installing the [fp8] extra if your hardware supports it, saving substantial system RAM during offloading.
• Run benchmarks with examples/benchmark.py to measure end-to-end pipeline stage times and peak memory usage on your specific hardware configuration.
• For multi-GPU setups, launch the pipeline with torchrun and initialize the process group to shard execution across available devices.
• Review the license terms carefully, especially if you are operating in restricted regions, to ensure compliance with the MiniMax H3 Community License Agreement.