HyperFlow H3 Flash: Compressed MiniMax-H3 Video and Audio Generation

Overview

HyperFlow H3 Flash is a high-performance, compressed version of the MiniMax-H3 video and audio generation model. By reducing the original ~31B parameter transformer to approximately 14B parameters, the model achieves efficient inference in 4 to 8 steps without classifier-free guidance (CFG). It supports text-to-video, image-to-video, and reference-guided generation with synchronized 32 kHz stereo audio.

Performance and Hardware Requirements

The model is optimized for both consumer RTX GPUs and data-center hardware. Performance varies based on the chosen runtime and available system memory:

  • GGUF Runtime: Recommended for consumer PCs with 16 GB to 64 GB of system RAM. It is compatible with ComfyUI, Unsloth Studio, and stable-diffusion.cpp. The Q4_K_M quantization requires approximately 9.1 GB of VRAM.
  • INT8 Runtime: Designed for data-center GPUs (H100/H200/B200) or high-end RTX GPUs (24 GB+ VRAM). This runtime utilizes W8A8 INT8 quantization with low-rank factorization and requires significant system RAM (69 GB to 91 GB depending on text encoder precision).

On 4× Blackwell GPUs, the model can achieve real-time inference. On a single RTX 5090, a 5.2-second clip can be generated in approximately 92 seconds using the INT8 runtime.

Architecture and Optimization

HyperFlow H3 Flash achieves its footprint reduction through three primary architectural modifications:

  1. Structural Pruning: Selected transformer blocks are removed, and linear layers are replaced with low-rank factorized versions to reduce parameter count to 13.7B–13.9B.
  2. Rotated W8A8 INT8 Quantization: 180 linear layers are quantized using a block-diagonal Hadamard rotation to maintain precision.
  3. Two-Time (t, r) Conditioning: The model uses a dual-conditioning approach where each token receives the current noise level (t) and the step endpoint (r), allowing for high-quality generation in as few as 4 steps.

Deployment and Integration

Users can deploy the model via two main paths:

  • ComfyUI: The project provides custom nodes and templates for text-to-video (t2va), image-to-video (fl2va), and reference-guided (ref2va) workflows. It is recommended to use the --fast-disk flag on systems with 32 GB or less of RAM to stream weights from the page cache.
  • Python CLI: The hyperflow-h3-flash package supports multi-GPU execution using torchrun. It implements Ulysses context parallelism to split sequences across GPUs and performs parallel VAE tile decoding to maintain bitwise consistency with single-GPU outputs.

Community Insights

Users have noted the utility of the model for local video generation, though some have expressed interest in more detailed comparative benchmarks between the various quantization levels. While ComfyUI is the primary integration, the project also provides a CLI for users seeking a more streamlined, terminal-based workflow similar to llama.cpp.

Sources