MiniMax H3 tops a video-editing arena, but its open weights exclude US users

MiniMax released the core weights for its H3 video model, making it the first publicly downloadable model to lead a category in Artificial Analysis' video arena, ranking first for video editing, second for text-to-video, and third for image-to-video as of August 3, 2026. However, the 33-billion-parameter H3-Base weights exclude users in the US, EU, UK, and South Korea under its custom license, and the hosted prompt-processing and 2K regeneration modules were not released.

MiniMax H3 tops a video-editing arena, but its open weights exclude US users - H3 ranked first for video editing, second for text-to-video and third for image-to-video in Artificial Analysis results reported on August 3, 2026. 1 https://the-decoder.com/chinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking/ - MiniMax released the 33-billion-parameter H3-Base weights, but withheld the system’s hosted prompt-processing and 2K regeneration modules. 2 https://huggingface.co/MiniMaxAI/MiniMax-H3 - The community license does not authorize use of the weights or their outputs in the US, EU, UK or South Korea. 3 https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE MiniMax has released the core weights for its H3 video model, making it the first publicly downloadable model to lead a category in Artificial Analysis’ video arena. As reported on August 3, H3 ranked first for video editing, second for text-to-video and third for image-to-video. 1 https://the-decoder.com/chinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking/ That result is narrower than an overall victory against proprietary systems from ByteDance, Google or Kuaishou. Artificial Analysis maintains six separate ranking pools, dividing text generation, image animation and editing into audio and silent categories. Users choose between blinded outputs made from the same prompt, and the evaluator recalculates ratings hourly using a Bradley-Terry preference model. 4 https://artificialanalysis.ai/video/methodology Open weights with major omissions The release includes two H3-Base checkpoints for text and first-or-last-frame generation, and for multimodal reference generation. The core is a dense 33-billion-parameter transformer paired with Qwen3-VL-32B as its encoder. It produces four- to 15-second clips at 24 frames per second with stereo audio; local output has a 768-pixel short edge. 2 https://huggingface.co/MiniMaxAI/MiniMax-H3 MiniMax did not release H3-Context-IR, the hosted system that interprets and restructures complex prompts, or H3-Regenerate-2K, which turns base output into higher-resolution video. The complete Hugging Face repository is 498 GB, while MiniMax’s reference SGLang configuration uses four GPUs. ComfyUI offers a pruned, quantized transformer file for smaller systems, but generation speed and quality will depend heavily on hardware and resolution. 2 https://huggingface.co/MiniMaxAI/MiniMax-H3 5 https://docs.comfy.org/tutorials/video/minimax/minimax-h3 License and testing limit the claim Calling H3 open source would be misleading. Its custom license excludes the United States, European Union, United Kingdom and South Korea; companies exceeding $20 million in annual revenue also need separate commercial authorization. US developers can access MiniMax’s hosted API, but the published license does not authorize them to run the released weights locally. 3 https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE MiniMax says its hosted prompt workflow automatically screens text, images and videos for unlawful, pornographic or potentially rights-infringing material. Those controls are tied to hosted services and are not described as embedded safeguards in the downloadable base model; local operators instead inherit obligations under the acceptable-use policy. 6 https://huggingface.co/MiniMaxAI/MiniMax-H3 h3-context-ir The arena is useful because it compares outputs blindly under consistent settings, but it measures aggregate preference rather than reliability, copyright risk or instruction accuracy. Early third-party comparisons remain informal: one same-prompt test found H3 stronger on material detail while ByteDance’s Seedance produced smoother animation and transitions. Broader reproducible testing had not appeared by August 4. 7 https://www.reddit.com/r/Seedance AI/comments/1vf2o5x/minimax h3 vs seedance 20 same prompt different/ Companies mentioned Further sources 1 The Decoder, “China's MiniMax H3 is the first open model to top an AI video ran… ↗ https://the-decoder.com/chinas-minimax-h3-is-the-first-open-model-to-top-an-ai-video-ranking/ 2 MiniMax H3 official model card and architecture documentation on Hugging Face. ↗ https://huggingface.co/MiniMaxAI/MiniMax-H3 3 MiniMax H3 Community License Agreement, dated August 2, 2026. ↗ https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE 4 Artificial Analysis video-generation benchmarking methodology. ↗ https://artificialanalysis.ai/video/methodology 5 ComfyUI’s official MiniMax H3 workflow and model documentation. ↗ https://docs.comfy.org/tutorials/video/minimax/minimax-h3 6 MiniMax H3 model card, H3-Context-IR safety guardrails section. ↗ https://huggingface.co/MiniMaxAI/MiniMax-H3 h3-context-ir +1 more The stories that matter, in one email. Free — unsubscribe anytime.