llms.blog
MiniMax

MiniMax Restricts H3 Video Model Access in Western Markets, Citing Copyright Uncertainty

Chinese AI company MiniMax has released the open weights for its H3 video generation model but imposed licensing restrictions on users in the United States, European Union, United Kingdom, and South Korea. Developers in those regions must apply for a formal license before using the model, the Shanghai-based company said Monday. The 33-billion-parameter H3 model processes text, images, video, and audio in a single prompt and can generate 4- to 15-second clips with stereo sound. A single prompt a

axc
axc
2 min
MiniMax Restricts H3 Video Model Access in Western Markets, Citing Copyright Uncertainty

Chinese AI company MiniMax has released the open weights for its H3 video generation model but imposed licensing restrictions on users in the United States, European Union, United Kingdom, and South Korea. Developers in those regions must apply for a formal license before using the model, the Shanghai-based company said Monday.

The 33-billion-parameter H3 model processes text, images, video, and audio in a single prompt and can generate 4- to 15-second clips with stereo sound. A single prompt accepts up to nine reference images, three video clips, and three audio clips. Artificial Analysis ranked H3 first in Video Editing, second in Text-to-Video, and third in Image-to-Video benchmarks, making it the first open-weight model to lead a video generation ranking.

MiniMax attributed the regional carve-out to regulatory uncertainty and ongoing copyright-related legal proceedings. The company said it would review deployment scenarios and confirm compliance controls before authorizing usage in the restricted markets. API access through paid subscriptions remains available globally.

The release is not fully open. The 2K resolution module and H3-Context-IR, a component that translates prompts and reference material into a structured intermediate format, remain proprietary. Local use through ComfyUI is capped at 768p resolution. Users must handle context preparation themselves using MiniMax's published prompting guides. The license also limits commercial use to companies with under $20 million in annual revenue, a restriction that blocks larger enterprises from deploying the model without a separate agreement.

Takedowns on decensor adapters

The restrictions follow a separate controversy over fine-tuned versions of H3. Within days of the release, community members published LoRA adapters that removed the model's built-in content safety filters. MiniMax responded by issuing takedown notices against those projects on platforms including Hugging Face, with employees warning that failure to remove them would result in license revocation. Several decensor LoRAs have since been deleted.

The H3 launch comes on the same day ByteDance released its closed-weight Seedance 2.5 model, which generates 30-second clips with built-in audio. The contrast between the two companies reflects a growing divide in Chinese AI release strategies: MiniMax opted for conditional openness, while ByteDance kept its model entirely behind an API.

MiniMax has not specified what criteria it will use to approve or deny license applications from the restricted regions, nor has it provided a timeline for when those restrictions might be lifted.

Sources

South China Morning Post: China's MiniMax curbs overseas access to new AI video model over copyright disputes — https://www.scmp.com/tech/tech-trends/article/3362951/chinas-minimax-curbs-overseas-access-new-ai-video-model-over-copyright-disputes

The Decoder: China's MiniMax H3 is the first open model to top an AI video ranking — https://the-decoder.com/chinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking/

MiniMax H3 model card on Hugging Face — https://huggingface.co/MiniMaxAI/MiniMax-H3

MiniMaxOpen Sourcevideo generation
ShareX (Twitter)LinkedIn

Written by

axc

More to read

  • Embedding Model Fine-Tuning in Production: Hard Negative Mining, Synthetic Data Pipelines, and Contrastive Distillation

    Production retrieval-augmented generation (RAG) and semantic search architectures frequently suffer from domain mismatch when relying on general-purpose embedding models. Off-the-shelf bi-encoders trained on broad web corpora often experience a 15% to 30% degradation in retrieval metrics such as NDCG@10 and MRR@10 when deployed on specialized enterprise corpora, including proprietary codebases, internal API schemas, clinical trials, and technical documentation. While downstream cross-encoder re

    1 min

  • The Gumbel-Softmax Trick: How Continuous Relaxations Enable Differentiable Discrete Sampling

    The Gumbel-Softmax Trick: How Continuous Relaxations Enable Differentiable Discrete Sampling In modern deep learning, end-to-end training depends on reverse-mode automatic differentiation. When an architecture operates on continuous tensors, computing gradients via the chain rule is straightforward. However, many foundational artificial intelligence problems involve discrete choices: selecting tokens from a fixed vocabulary, routing tokens to expert networks in a Mixture-of-Experts (MoE) archit

    1 min

  • Token Healing and Partial Token Alignment in Production LLM Serving: Architecture, Prefix Trie Rollback, and Serving Trade-Offs

    Modern large language models operate on discrete subword tokens generated by greedy compression algorithms like Byte-Pair Encoding (BPE), WordPiece, or Unigram. While subword tokenization enables high compression rates and fixed vocabulary sizes, it introduces a structural defect at the interface between raw user text and autoregressive inference: the partial token problem, commonly known as the prompt boundary problem. When a user prompt terminates mid-token or at a punctuation boundary that c

    1 min