HyperFlow H3 Flash explained: 13.7B video model, 6 steps

By LLMTest Lab · Nov 7, 2026 · 4 min read videoreleaseopen-weights
On this page

On this page

  1. What is new versus MiniMax-H3
  2. Speed and hardware, as reported
  3. How it compares on paper
  4. Who should care

Not tested by us yet. We generate no video in our tests, so every number below comes from Video Rebirth's README and model card on Hugging Face, and all of it is vendor-reported. We'll update this post when independent side-by-side clips against full-size MiniMax-H3 appear.

HyperFlow H3 Flash is a compressed, few-step version of MiniMax-H3 that makes 5.2 to 15 second video clips with synchronized stereo audio. The transformer drops from about 31B to 13.7B parameters, sampling takes 4, 6 or 8 steps without classifier-free guidance, and the weights are public. If you already self-host video models, it is worth a download. If you want a hosted API, there is none yet.

HyperFlow H3 Flash
Maker Video Rebirth
Base model MiniMax-H3
Parameters 13.7B (fl2va), 13.9B (ref2va)
Output 5.2 to 15 s at 24 fps, 1344x768 by default, 32 kHz stereo audio
Sampling 4, 6 (default) or 8 steps, no CFG
Tasks Text to video+audio, first/last keyframe, up to 12 references
Weights Hugging Face, GGUF and INT8
License MiniMax-H3 Community License Agreement
Hosted API None announced

What is new versus MiniMax-H3

The README lists three changes to the MiniMax-H3 transformer. First, structural pruning and low-rank factorization: 3 of 50 blocks are dropped for text, keyframe and image-to-video work (1 of 50 for reference work), and 102 or 114 linear layers are factored. Second, 180 linear layers are quantized to W8A8 INT8 behind a Hadamard rotation. Third, a two-time conditioning scheme feeds every token both the current noise level and the step endpoint, which is what allows 4 to 8 step sampling without CFG.

The practical result is a model less than half the size of its parent that keeps joint video and audio generation. The same weights cover text to video, first and last keyframe video, and reference-to-video with up to 9 images, 3 videos and 3 audio clips.

Speed and hardware, as reported

Video Rebirth's landing page claims a 12.7x speedup over the H3 baseline on a single B200 and says output quality was checked against the bf16 baseline. Those are the maker's claims, and the page shows no side-by-side numbers we could verify.

The README benchmarks are more specific. All are end-to-end at 1344x768 and 6 steps, for a 5.2 s clip:

Hardware 5.2 s clip 15 s clip
4x B200 8.7 s not listed
4x H200 12.6 s 66 s
1x H200 36 s 229 s
RTX 5090 (GGUF) 99 s 575 s
RTX 4090 (GGUF) 157 s 830 s
RTX 3090 (GGUF) 315 s 1835 s

The consumer rows are the notable part. The Q4_K_M GGUF is 9.1 GB, and Video Rebirth says it runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM. The INT8 path is 10 to 20% faster but needs at least 69 GiB of system RAM, and the first-run download is about 90 GB. A 16 GB GPU is not measured yet.

How it compares on paper

The closest comparison is its own parent. MiniMax-H3's transformer is the roughly 31B model HyperFlow was cut down from, and the README says the 62 GB base transformer is never downloaded. Pika has also released an open model built on MiniMax-H3, Pika H3 Flex. For hosted video pricing, see our Gemini Omni 1.2 Flash explainer. A like-for-like comparison would need the same prompts through all three, and nobody has published one.

Treat "real time" carefully. The landing page ties it to four B200 GPUs at 768p. On a single RTX 5090, a 5.2 s clip takes about 99 s in the vendor's GGUF numbers, roughly 19 times the clip length.

Who should care

  • Teams that self-host video generation and have a 24 GB or larger GPU. This is the audience the GGUF and ComfyUI nodes target.
  • Anyone building on MiniMax-H3 who wants fewer steps. An 8-step LoRA for the full-size model is also linked from the README.
  • Anyone who needs an API. Not yet. The license also carries MiniMax's community terms, so read them before commercial use.
  • Anyone choosing on quality. Wait for independent clips. "Quality maintained" is a vendor sentence, not a result.

For how to turn a unit price into a per-task budget once a hosted option exists, see our post on LLM cost per feature. For another small open release from the same week, read about Underdog Woof, a 32 MB text-to-speech model.

Related articles

Gemini Omni 1.2 Flash explained: 15 s clips, 30% cheaper
Gemini Omni 1.2 Flash makes video with audio up to 15 seconds at about $0.07 per second in 720p, 30% below Omni 1.1. Specs, price per clip, who should switch.
Underdog Woof first look: a 32MB open TTS model on CPU
Underdog Woof is a 38M-parameter English text-to-speech model that runs on CPU in a 31.7 MB download. Specs, vendor-reported WER, license limits, who it suits.