Video-Rebirth/hyperflow-h3-flash

What it is

HyperFlow H3 Flash is a compressed derivative of MiniMax-H3 that produces 5.2–15 s of 24 fps video together with a 32 kHz stereo soundtrack. It samples in 4, 6 (default) or 8 steps without classifier-free guidance. Compared with the ~31B MiniMax-H3 transformer, the model has 13.7B parameters (fl2va) or 13.9B (ref2va) after pruning three or one of 50 blocks respectively and factoring 102/114 linear layers into low-rank form.

Two integer runtime formats are provided: rotated W8A8 INT8 (convrot) for data-center and 24 GB+ RTX GPUs, and GGUF Q3_K_M–Q8_0 for consumer GPUs through ComfyUI, Unsloth Studio or stable-diffusion.cpp. A two-time (t, r) conditioning scheme lets each token see both the current noise level and the step endpoint, which is what enables few-step sampling without CFG.

Three tasks are supported: t2va (text only), fl2va (prompt plus first keyframe, or first and last keyframes) and ref2va (prompt plus up to 12 references). Under INT8, ref2va accepts images, videos and audio; under GGUF it accepts images only.

Quick start with the Python CLI

Requirements are Python 3.10+, PyTorch 2.6+ with CUDA and Triton, an Ampere or newer NVIDIA GPU, and diffusers >= 0.40. Install the package and, on H100/H200, the FlashAttention-3 extra:

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

The first run downloads roughly 90 GB of weights (about 68 GB with --text-encoder-precision fp8) and compiles Triton INT8 kernels for about a minute. Then generate a clip:

hyperflow-h3-flash --prompt "A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window." \
    --output ramen.mp4

The result is an MP4 containing video and audio. Useful flags include --steps 4|6|8, --num-frames (124–362, rounded up to 17n + 5), --reference for ref2va, --gpus for context parallelism, and --offload for memory placement. The Python API exposes the same pipeline through load_pipeline("t2va") and returns videos, audio and the sampling rate for encoding.

For consumer hardware, the GGUF setup guide covers ComfyUI, Unsloth Studio and stable-diffusion.cpp, using the ComfyUI-HyperFlow-Flash nodes and Q4_K_M as the recommended default for 24 GB GPUs. Weights live in the Hugging Face repository.

Memory and performance constraints

The two runtimes target different hardware. GGUF with Q4_K_M is 9.1–9.2 GB and runs on 24/32 GB RTX cards with at least 16 GB of system RAM; on 16 GB and 32 GB machines, ComfyUI should be started with --fast-disk to stream weights from NVMe. INT8 is faster by 10–20% on the transformer but needs more host memory: about 69 GiB with the fp8 text encoder or 91 GiB with bf16 when offloading. Multi-GPU (--gpus > 1) requires Linux or WSL2. On 24 GB GPUs, 15 s mixed ref2va runs need --block-chunk 16384.

The repository's own benchmarks, at 1344×768 and 6 steps including text encoding and both VAE decodes, report 92 s (5.2 s clip) and 551 s (15 s clip) on an RTX 5090 with INT8, and 8.7 s on 4× B200. These are the authors' measurements, not independent verification. Real-time inference on 4× B-series GPUs is routed to NVIDIA SOL-H3.

Licensing limits

Repository code is Apache-2.0, with the bundled ComfyUI nodes under GPL-3.0 and the stable-diffusion.cpp patch under MIT. The weights are a Model Derivative of MiniMax-H3 under the MiniMax H3 Community License; downloading and using them means accepting that agreement. The agreement does not cover use in the European Union, the United Kingdom, the Republic of Korea or the United States without MiniMax's authorization, and includes an Acceptable Use Policy.

HotRepo's take

The project appeared on the HotRepo new-project list for Python on 2026-11-08, gaining 11 stars in a roughly 5.5-hour window and one fork. Its Show HN thread from 2026-11-07 drew 66 points and 13 comments, and the main criticism there was that the release lacks quality comparisons between quantizations and against the original model. A related request asked for a simpler CLI-like path instead of ComfyUI. The repository does provide both INT8 and GGUF routes and a CLI, but the quality question remains open.