Video Example

https://www.youtube.com/watch?v=zUk6mqruWdA


MiniMax-H3 Merged Dual Turbo LoRAs

This repository hosts 2 custom merged LoRA adapters designed for the MiniMax-H3 architecture. These adapters combine ultra-fast few-step inference, high-fidelity temporal consistency, synchronized stereo audio-video generation, and responsive streaming capabilities by merging three state-of-the-art community LoRA checkpoints:

  1. larryvrh/MiniMax-H3-Turbo-Lora (Apache-2.0)
  2. TaoLiveAIGC/TaoMate-H3 (MiniMax H3 Community License Agreement)
  3. ZhengmingYu/DMAD (minimax_h3) (MiniMax H3 Community License Agreement / DMAD Adversarial Distillation)

Models

  • Variant 1: turbo_lora_001.safetensors
  • Variant 2: turbo_lora_002.safetensors

Merge Details & Methodology

  • Base Foundation: MiniMaxAI/MiniMax-H3
  • Upstream Adapters Merged:
    • larryvrh/MiniMax-H3-Turbo-Lora (Few-step joint audio-video Turbo LoRA)
    • TaoLiveAIGC/TaoMate-H3 (Rank-128 low-latency streaming audio-video EMA LoRA)
    • ZhengmingYu/DMAD (minimax_h3/ Rank-128 DMAD student LoRA)
  • Merge Strategy: Linear weight interpolation (SLERP / DARE-TIES / Weighted Add) applied across matching attention and feed-forward projection layers.
  • Precision: bfloat16 / fp16

Inference Guidelines

  • Recommended Sampling Steps: 6 (Range: 5 - 9)
  • Guidance Scale (CFG): 1.0 (Recommended for native CFG-distilled operation)
  • Scheduler / Sampler: lcm + simple
  • Base Model Loading: Load on top of standard MiniMaxH3Transformer3DModel (compatible with Diffusers and ComfyUI).

Legal Notice, Terms & Licensing

1. Primary License & Mandatory Copyright Notice

Because these checkpoints are Model Derivatives of MiniMaxAI/MiniMax-H3 and incorporate weights from TaoLiveAIGC/TaoMate-H3 and ZhengmingYu/DMAD (minimax_h3/), this repository is distributed under the MiniMax H3 Community License Agreement.

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

By downloading, copying, modifying, distributing, or using these weights, you agree to comply with all terms and use restrictions set forth in the MiniMax H3 Community License Agreement and its Acceptable Use Policy (Exhibit A).

2. Geographic Restrictions (Excluded Territories)

IMPORTANT: In strict accordance with Section V.4 and Definitions 3 & 5 of the MiniMax H3 Community License Agreement, you may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory (i.e., within the Excluded Territories), specifically:

  • The United States of America (US)
  • The European Union (EU)
  • The United Kingdom (UK)
  • The Republic of Korea (South Korea)

Any use within the Excluded Territories is not authorized by the Community License. Users or organizations located in or operating services within these regions must obtain separate written authorization or commercial waivers directly from MiniMax at https://platform.minimax.io/h3-license (or via https://platform.minimax.io/console/license).

3. Commercial Caps, Attribution & Prohibited Uses

  • Annual Revenue Limit ($20M USD Ceiling): Commercial use within the Applicable Territory is permitted under the Community License provided the commercial products or services offered by the licensee do not generate annual gross revenue exceeding US$20,000,000 ($20M USD). Entities exceeding this threshold require separate written commercial authorization from MiniMax.
  • Hosted Services & UI Attribution: Distributing this model as a managed API or third-party service requires compliance safeguards and binding downstream users to terms at least as protective as the MiniMax H3 Community License Agreement. Any commercial product or service using these weights must prominently display "MiniMax H3" in its user interface and documentation.
  • Acceptable Use Policy & AI Labelling: Users must comply with the MiniMax H3 Acceptable Use Policy (Exhibit A) and clearly disclose that generated audio-video outputs are machine-generated.
  • No Competing Training: Neither these weights nor any Outputs generated by them may be used to train, distill, or improve any competing foundational AI models (other than MiniMax-H3 derivatives).

4. Upstream Model Derivatives & Apache 2.0 Notice of Modifications

This repository merges weights from two license categories:

  1. MiniMax H3 Community License Derivatives:
    • TaoLiveAIGC/TaoMate-H3 — Copyright © 2026 Alibaba TaoLive AIGC Team (Qijun Gan et al.). Licensed under the MiniMax H3 Community License Agreement.
    • ZhengmingYu/DMAD (minimax_h3/) — Copyright © 2026 Zhengming Yu, Junkun Yuan, Haotian Yang, Gordon Guocheng Qian, Chongyang Ma et al. Distributed as Model Derivatives of MiniMax H3 under the MiniMax H3 Community License Agreement.
  2. Apache License 2.0 Derivative Component (Section 4 Notice):

Statement of Modifications: In compliance with Section 4(b) of the Apache License 2.0 and the MiniMax H3 Community License Agreement, notice is hereby given that the original LoRA weights from larryvrh/MiniMax-H3-Turbo-Lora, TaoLiveAIGC/TaoMate-H3, and ZhengmingYu/DMAD (minimax_h3/) were modified by mathematical layer re-weighting, key standardization, and linear interpolation into two new composite .safetensors adapter files (turbo_lora_001.safetensors and turbo_lora_002.safetensors).

5. Disclaimer of Warranty

THESE WEIGHTS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED. The author of this merge is an independent developer and is not affiliated with or endorsed by MiniMax (Nanonoble Pte. Ltd.), Alibaba TaoLive AIGC Team, larryvrh, or the DMAD research authors.


Citations & Acknowledgments

Credit goes to the original authors and research teams:

@misc{yu2026dmad,
  title={DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation},
  author={Yu, Zhengming and Yuan, Junkun and Yang, Haotian and Qian, Gordon Guocheng and Wang, Yizhi and Wang, Angtian and Yang, Yiding and Liu, Bo and Li, Xin and Wang, Wenping and Ma, Chongyang},
  year={2026},
  eprint={2610.02188},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={[https://arxiv.org/abs/2610.02188](https://arxiv.org/abs/2610.02188)}
}

@misc{gan2026taomate,
  title={TaoMate: Anchor-Guided Memory Bridging Evolving and Reference States for Real-Time Audio-Video Digital Human Generation},
  author={Gan, Qijun and Zhang, Chenwei and Jin, Meiguang and Ma, Junfeng and Shen, Qiu},
  year={2026},
  eprint={2607.24359},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={[https://huggingface.co/TaoLiveAIGC/TaoMate-H3](https://huggingface.co/TaoLiveAIGC/TaoMate-H3)}
}

@misc{larryvrh2026turbo,
  title={MiniMax-H3-Turbo-Lora},
  author={larryvrh},
  year={2026},
  url={[https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora)}
}

@misc{minimax2026h3,
  title={MiniMax-H3 Multimodal Generative Model},
  author={{MiniMax AI}},
  year={2026},
  url={[https://huggingface.co/MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3)}
}
Downloads last month
-
Inference Providers NEW

This task can take several minutes

Model tree for Raretutor/MiniMax-H3-Dual-Turbo-Lora-ComfyUI

Adapter
(203)
this model

Papers for Raretutor/MiniMax-H3-Dual-Turbo-Lora-ComfyUI