Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

HyperFlow H3

Video Rebirth's few-step versions of MiniMax H3: a top-five arena score, a 14B Flash that fits a 24 GB card, and H3's restrictive licence.

Reviewed by ReviewRamble editors · Updated

Visit HyperFlow H3 Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

HyperFlow H3 makes MiniMax's video model fast enough to run at home, as long as home is not in the four markets its licence excludes.

What is HyperFlow H3?

HyperFlow H3 is Video Rebirth's family of faster versions of MiniMax H3, made by what the company calls data-free flow self-distillation: the model generates candidate videos, filters them for quality and learns from the ones that pass, with no outside data or human labels. The first release, in September 2026, is an 8-step LoRA adapter of 2.8 GB for the full MiniMax H3 transformer; where the standard diffusers schedule runs 49 model passes, HyperFlow runs 8. It covers text-to-video, first and last frame, and multi-reference generation, all with sound.

HyperFlow H3 Flash followed on November 6, 2026: a pruned model of 13.7 billion parameters (13.9 billion for the reference variant), less than half the size of the roughly 31-billion-parameter H3 transformer. It samples in 4, 6 or 8 steps without classifier-free guidance and outputs 5.2 to 15 second clips at 24 fps, 1344 by 768 by default, with 32 kHz stereo sound. Video Rebirth reports a 5.2-second clip in 8.7 seconds on four B200 GPUs, 36 seconds on one H200 and 92 seconds on an RTX 5090, and its 4-bit GGUF file is 9.1 GB.

The inference code is Apache 2.0, but the weights inherit the MiniMax H3 Community License, which does not license use in the EU, UK, South Korea or the US without MiniMax's prior authorisation. Video Rebirth also sells the model through its own API.

Is HyperFlow H3 worth it?

HyperFlow H3 is the most practical way to run MiniMax H3 yourself, and the licence it inherits is the reason most Western buyers cannot. On Artificial Analysis's text-to-video board the 8-step version ranks 5th at 1155 Elo, 7 points behind H3 itself at 1162, and on image-to-video it is 8th at 1121. Cutting 49 model passes to 8 cost almost nothing in the blind votes.

The Flash model is the more interesting release. At 13.7 billion parameters and 9.1 GB in its 4-bit form, it runs on 24 GB and 32 GB graphics cards, and Video Rebirth times a 5.2-second clip at 92 seconds on an RTX 5090. Video Rebirth has not published arena scores for Flash, and in ComfyUI Kijai measured only about a 10% real-world speedup, well short of what the step count suggests.

The catch: the MiniMax H3 Community License does not cover the EU, UK, South Korea or the US without MiniMax's permission, and HyperFlow cannot loosen it. The hosted route is no bargain either: Artificial Analysis lists Video Rebirth's API at $9.00 a minute for image-to-video, nearly twice H3's own $4.80, and slightly below H3 on that board.

Our take: HyperFlow is the version of H3 for people who own their GPUs and live outside the licence's exclusions. Everyone else gets more from MiniMax's own API, which is cheaper, or from Wan 3.0, which scores higher.

Where it earns its place

  • A solo motion designer in Singapore with an RTX 5090

    Generate short clips with sound locally, without paying per second

    Load the 9.1 GB Q4_K_M Flash GGUF in ComfyUI with the HyperFlow templates and render 5-second drafts in about a minute and a half each. Cost is electricity, but the licence still bars selling output into the excluded markets.

  • An ML team at a Southeast Asian video app already serving MiniMax H3

    Cut inference cost on existing H3 workflows

    Swap in the 2.8 GB HyperFlow LoRA on the official diffusers pipeline to run 8 model passes instead of 49, keeping the same text-to-video, first-and-last-frame and reference workflows.

Before you commit

Setup
Download the LoRA or the Flash weights from Hugging Face, install the hyperflow or hyperflow-h3-flash package or the ComfyUI nodes, and fetch H3's text encoder and VAEs separately.
Integrations
diffusers Modular Pipeline, ComfyUI templates, stable-diffusion.cpp, Unsloth Studio; NVIDIA's SOL-H3 engine supports Flash; Video Rebirth API for hosted use.
Data and privacy
Self-hosted runs keep prompts and footage on your own hardware; Video Rebirth's API retention terms were not published on the pages we read.
Cost at scale
Self-hosted: GPU time only. Hosted: at Artificial Analysis's $9.00 a minute, 100 minutes of image-to-video a month is $900.
Limits
Clips of 5.2 to 15 seconds at 768p by default; 16 GB GPUs not measured; Flash reference inputs in GGUF are images only; licence excludes the EU, UK, South Korea and the US.

What’s good

  • 8-step LoRA ranks 5th on Artificial Analysis's text-to-video board
  • 14B Flash model runs on a 24 GB GPU in ComfyUI, with stereo sound
  • Inference code under Apache 2.0, with GGUF and INT8 builds

What’s not

  • Weights not licensed in the EU, UK, South Korea or the US without MiniMax's authorisation
  • Hosted API costs nearly twice MiniMax's own H3 per minute
  • No independent scores for Flash; real-world ComfyUI speedup about 10%

Best for

Studios and hobbyists outside the excluded markets who want H3-quality video with sound on their own GPUs.

HyperFlow H3 alternatives

HyperFlow H3: common questions

Is HyperFlow H3 worth it?
HyperFlow H3 makes MiniMax's video model fast enough to run at home, as long as home is not in the four markets its licence excludes. We rate it 3.9 out of 5.
How much does HyperFlow H3 cost?
Weights free under the MiniMax H3 Community License. Video Rebirth API: Artificial Analysis lists $9.00 per minute for image-to-video and $18.00 for video editing at 1088p.
What are the downsides of HyperFlow H3?
Weights not licensed in the EU, UK, South Korea or the US without MiniMax's authorisation; hosted API costs nearly twice MiniMax's own H3 per minute; no independent scores for Flash; real-world ComfyUI speedup about 10%.
Who is HyperFlow H3 best for?
Studios and hobbyists outside the excluded markets who want H3-quality video with sound on their own GPUs.
How much does HyperFlow H3 cost for a team or in production?
Self-hosted: GPU time only. Hosted: at Artificial Analysis's $9.00 a minute, 100 minutes of image-to-video a month is $900.
Is HyperFlow H3 safe for company data?
Self-hosted runs keep prompts and footage on your own hardware; Video Rebirth's API retention terms were not published on the pages we read.
What are the best HyperFlow H3 alternatives?
The closest alternatives we have reviewed are MiniMax H3 (Hailuo 3.0), Pika H3 Flex, LTX-2.5, Wan 3.0 and HeyGen Video 1.

Answers checked Nov 21, 2026.