EditorNodesPricingBlog

HyperFlow H3 Flash: Open Source 4-Step Video LoRA & GGUF

November 15, 2026
HyperFlow H3 Flash: Open Source 4-Step Video LoRA & GGUF

Video Rebirth, HyperFlow H3 Flash (Apache 2.0 Code / MiniMax H3 Community License)

Share this post:

HyperFlow H3 Flash: Open Source 4-Step Video LoRA & GGUF

HyperFlow H3 Flash (VideoRebirthAI/hyperflow-h3-flash and Video-Rebirth/hyperflow-h3-flash) is an open source, structurally pruned 13.7B to 13.9B video and stereo audio generation suite developed by Video Rebirth on top of the ~31B MiniMax-H3 architecture. Released alongside the 2.8 GB rank 256 HyperFlow 8-Step LoRA (VideoRebirthAI/hyperflow), HyperFlow H3 Flash compresses the base transformer by more than half and executes joint 1344x768 24 fps video and 32 kHz stereo audio generation in 4, 6, or 8 steps (CFG=1.0) using GGUF (Q3_K_M to Q8_0) and rotated W8A8 INT8 safetensors (convrot). The inference code is licensed under Apache 2.0, while the model weights follow the MiniMax H3 Community License Agreement.

Studio availability. HyperFlow H3 Flash is not available in AI FILMS Studio. This article covers a public research release. To generate video in the Studio, see the models listed on the Image to Video page or in Studio Nodes.

HyperFlow H3 Flash 6 step 768p video and synchronized stereo audio generation

How HyperFlow H3 Flash Cuts Latency and VRAM in Half

Running the stock 31B MiniMax-H3 transformer in Hugging Face diffusers requires 49 neural function evaluations (50 sigma points) and multi GPU data center memory. Video Rebirth addresses both the step count and the memory footprint through three technical stages.

  1. Two time flow map conditioning (t, r). Built on the AnyFlow formulation, a TwoTimeEmbedder wraps the base time_embedder and adds a second embedding for the step endpoint r, blended via a learned gate stored in the safetensors header. Conditioning tokens (fl2va keyframes and ref2va reference media) keep r = t so reference frames remain pinned while the denoiser jumps across a 4, 6, or 8 step sigma grid (video_shift = 12, audio_shift = 3) with no classifier free guidance overhead (CFG = 1.0).
  2. Structural pruning and low rank factorization. Rather than uniformly shrinking layer widths, the fl2va partition drops 3 of 50 transformer blocks (blocks 6, 21, and 22) and factors 102 linear layers to reach 13.7B parameters. The ref2va partition drops block 9 and factors 114 linear layers to reach 13.9B parameters, preserving multi modal reference routing for up to 12 inputs (up to 9 images, 3 videos, and 3 audio tracks).
  3. Rotated W8A8 INT8 (convrot) and GGUF quantization. For fast tensor core execution, 180 linear layers are quantized to W8A8 INT8 behind a block diagonal Hadamard rotation (~16 GB), dispatching to fused Triton or cuBLASLt INT8 kernels. For lower memory footprints, GGUF quantizations shrink the transformer down to 9.1 GB (Q4_K_M) or 7.2 GB (Q3_K_M).
HyperFlow H3 Flash 768p real time generation preview frame

HyperFlow H3 Flash renders 5.2s to 15s 768p clips with 32 kHz stereo soundtracks in 8.7s on 4x B200 GPUs. Credit: Video Rebirth.

HyperFlow H3 Flash running on consumer RTX GPUs via INT8 and GGUF

With 9.1 GB Q4_K_M GGUF and 16 GB INT8 convrot weights, the 14B pipeline runs locally on 24 GB RTX 4090 and 32 GB RTX 5090 cards. Credit: Video Rebirth.

HyperFlow H3 Flash 6 step output preserving camera motion, lighting detail, and stereo audio synchronization

Weight Formats, Hardware & Benchmark Matrix

All benchmarks below measure complete generation (text encoding, 6 step denoising at 1344x768, and video plus stereo audio VAE decoding) for a 5.2 second clip (124 frames at 24 fps) and a 15 second clip (362 frames).

Weight Variant / Adapter Hugging Face Path Size & Precision Hardware & Total Speed (5.2s / 15s) License
HyperFlow H3 Flash INT8 VideoRebirthAI/hyperflow-h3-flash (fl2va_int8_convrot/, ref2va_int8_convrot/) ~16.0 GB (W8A8 INT8 + bf16 low rank) 4x B200 (8.7s), 1x H200 (36s / 229s), RTX 5090 32 GB (83s / 518s), RTX 4090 24 GB (130s / 750s) MiniMax H3 Community License
HyperFlow H3 Flash GGUF Q4_K_M VideoRebirthAI/hyperflow-h3-flash (fl2va_gguf/, ref2va_gguf/) 9.1 GB (fl2va) / 9.2 GB (ref2va) RTX 5090 32 GB (99s / 575s), RTX 4090 24 GB (151s / 812s with --fast-disk on 16 GB RAM PC) MiniMax H3 Community License
HyperFlow H3 Flash GGUF Q8_0 VideoRebirthAI/hyperflow-h3-flash (fl2va_gguf/, ref2va_gguf/) 14.9 GB (fl2va) / 15.2 GB (ref2va) 24 to 32 GB VRAM GPUs for near full precision GGUF execution MiniMax H3 Community License
HyperFlow H3 Flash GGUF Q3_K_M VideoRebirthAI/hyperflow-h3-flash (fl2va_gguf/, ref2va_gguf/) 7.2 GB (fl2va) / 7.4 GB (ref2va) Smallest footprint for compact local setups (stable-diffusion.cpp / Unsloth Studio) MiniMax H3 Community License
HyperFlow 8-Step LoRA (31B) VideoRebirthAI/hyperflow (minimax_h3_hyperflow_8step_v1.0.safetensors) 2.8 GB (rank 256 PEFT LoRA for full 31B MiniMax-H3) 4x H200 (60s vs 175s baseline), 1x H200 (130s vs 395s baseline, ~3x speedup) MiniMax H3 Community License

Running HyperFlow H3 Flash Locally in Python and CLI

Install the hyperflow-h3-flash package with fp8 text encoder support (which saves ~22 GiB of system RAM during CPU offload) or Hopper FlashAttention-3 support.

pip install "hyperflow-h3-flash[fp8] @ git+https://github.com/Video-Rebirth/hyperflow-h3-flash"

# 1. Text to video + stereo audio (t2va) in 6 steps
hyperflow-h3-flash \
  --prompt "A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window." \
  --steps 6 \
  --text-encoder-precision fp8 \
  --output ramen_sync.mp4

# 2. First and last keyframe conditioning (fl2va)
hyperflow-h3-flash \
  --prompt "Camera tracks forward as the lantern lights up the misty alley." \
  --image first_frame.png \
  --last-image last_frame.png \
  --steps 6 \
  --output lantern_transition.mp4

# 3. Multi-modal reference to video + audio (ref2va)
hyperflow-h3-flash \
  --prompt "<Picture 1> walks through the neon street of <Picture 2> while speaking." \
  --reference actor.png \
  --reference street_plate.png \
  --steps 6 \
  --block-chunk 16384 \
  --output ref2va_scene.mp4

You can also invoke load_pipeline in Python with diffusers >= 0.40 or run the Q4_K_M GGUF checkpoint via stable-diffusion.cpp (sd-cli -M vid_gen).

import torch
from diffusers.utils.export_utils import encode_video
from hyperflow_h3_flash import load_pipeline

pipe = load_pipeline("fl2va")
out = pipe(
    prompt="A red fox trotting through a snowy pine forest with crisp footsteps in the snow.",
    image="first_keyframe.png",
    num_frames=124,
    num_inference_steps=6,
    generator=torch.Generator().manual_seed(42),
    output=["videos", "audio", "sampling_rate"],
)

encode_video(
    out["videos"][0],
    fps=24,
    output_path="fox_snow_6step.mp4",
    audio=out["audio"][0],
    audio_sample_rate=out["sampling_rate"],
)

For related MiniMax-H3 and GGUF acceleration workflows, see our guide to FreeVideo, MiniMax-H3, and Pika H3-Flex, our MiniMax-H3 video generation tutorial, and our walkthrough of Wan 2.2 GGUF 8 GB VRAM local workflows.

Fast Image-to-Video and Audio Workflows in AI FILMS Studio

HyperFlow H3 Flash runs locally on consumer and workstation GPUs and is governed by the territorial and commercial terms of the MiniMax H3 Community License Agreement. To generate high resolution image to video, keyframe transitions, and synchronized audio directly in the browser, open the video workspace or chain multi reference shots in the Nodes Graph Editor.


Sources