AI & Creative Tools

HyperFlow H3 Flash Promises a 12x Speedup for MiniMax H3 Video — the Community Measured 10%

Video Rebirth's compressed, few-step version of MiniMax H3 pruned three transformer blocks and quantized the rest to run video generation in 4-8 steps. The marketing says 12.7x faster; a widely-used ComfyUI contributor's own benchmark says otherwise.

Video generation models keep getting bigger, and the tools built to make them faster keep making bigger claims. The latest is HyperFlow H3 Flash, released November 6 by Video Rebirth, a compressed, few-step variant of MiniMax’s H3 video model. The pitch is a 12.7x speedup over the full model on a single NVIDIA B200 GPU. The reality, according to one of the most respected independent voices in the ComfyUI ecosystem, is closer to 10%.

What it actually does

MiniMax H3’s own transformer runs roughly 31 billion parameters across 50 blocks. HyperFlow H3 Flash gets there differently: it drops 3 of those 50 blocks (blocks 6, 21, and 22) and factors 102 of the remaining linear layers using low-rank decomposition, shrinking the effective parameter count to about 13.7 billion. On top of that, 180 linear layers get quantized to rotated W8A8 INT8 (“convrot” in the project’s own terminology). A technique the team calls two-time (t, r) conditioning is what lets the model sample in as few as 4 steps without classifier-free guidance — a meaningful change from how most diffusion-style video models need many more steps and a negative prompt to converge on something coherent.

Crucially, as Video Rebirth’s own documentation puts it, the architecture isn’t new: the speed comes from compressing an existing model, not designing a smaller one from scratch. That distinction matters for anyone trying to judge the output quality — a distilled model inherits the base model’s strengths and failure modes, for better or worse. The repository stores only the compressed transformer weights; the text encoder and video/audio VAEs still come from the upstream MiniMax H3 repository on Hugging Face.

The gap between the marketing and the measurement

Here’s where it gets interesting. Video Rebirth’s release materials lead with a 12.7x speedup over the H3 baseline, measured on a single B200. That is an eye-catching number for a tool aimed at people running video generation on a single GPU at home.

But Kijai — the developer behind some of the most widely-installed custom node packs in the ComfyUI ecosystem, whose benchmarks carry real weight in that community — measured something very different: roughly a 10% speedup over the already-pruned full model when run inside ComfyUI. Another contributor, LukeG89, was blunter, calling the tradeoff “not worth it.” Real-world timing figures from the model’s own card bear this out more than the headline number does: a 5.2-second clip at 1344×768 resolution takes about 92 seconds on a consumer RTX 5090, and a 15-second clip runs 500-575 seconds on that same card — fast for the quality on offer, but nowhere near the multiplier in the pitch.

None of this means HyperFlow H3 Flash is useless. A model that runs natively via GGUF quantization on consumer GPUs with 24GB of VRAM or less is still meaningfully more accessible than the full 31-billion-parameter original, and the benchmark gap doesn’t erase the convenience of 4-to-8-step sampling. What it does mean is that “12.7x faster” was Video Rebirth’s number, measured on Video Rebirth’s hardware, under Video Rebirth’s conditions — and the moment an independent, technically credible party reran the test on different hardware, the gap narrowed to something far less dramatic.

A restrictive license wrapped in an open-weights release

The weights are published on Hugging Face, but calling this “open source” would overstate things. The license is the MiniMax H3 Community License Agreement, which explicitly excludes the European Union, the United Kingdom, South Korea, and the United States from its default grant — users in those territories need MiniMax’s prior written authorization to use it at all. Products built on top of it that generate more than $20 million in annual revenue need separate authorization, and any product using the weights must display “MiniMax H3” prominently in its interface. Outputs can’t be used to train or improve other AI models, and any publicly distributed content made with it must be disclosed as AI-generated.

This is a familiar shape for 2026’s open-weight AI landscape: genuinely useful weights, published for anyone to download and run, wrapped in licensing terms that look more like a negotiated commercial agreement than a permissive open-source license. For anyone building a ComfyUI workflow around it, the practical lesson is the same one Kijai’s benchmark already taught: read the independent numbers before the marketing copy, on both speed and licensing.