MiniMax H3Open Weights

MiniMax H3 (also known as Hailuo 3.0) is a 33.1-billion parameter omni-modal video generation model developed by MiniMax. It is distinguished by its ability to jointly understand and generate text, images, video, and audio within a single, unified context. Unlike models that treat audio as a separate post-processing step, H3 generates high-fidelity video with synchronized native stereo audio—including dialogue, ambient sound, and foley—in a single inference pass.

The model is built on a dense single-stream transformer architecture (H3-Omni-Transformer), utilizing a Qwen3-VL-32B encoder backbone for multimodal reasoning. It employs three-dimensional Multimodal Rotary Position Embeddings (MM-RoPE) to represent complex spatial and temporal relationships across frames. This integrated design allows H3 to maintain high temporal consistency and physics-aware motion over durations of up to 15 seconds at a cinematic 24 frames per second, with official support for native 2K resolution.

H3 supports several advanced conditioning modes for precise creative control, including T2VA (Text-to-Video-Audio), FL2VA (First-and-Last-Frame-to-Video-Audio) for animating between two specific states, and Ref2VA (Reference-to-Video-Audio). In omni-reference mode, the model can process up to 12 total files simultaneously—up to 9 images, 3 video clips, and 3 audio tracks—to guide character identity, environmental style, and vocal performance. It currently provides stable support for 11 languages, including English, Chinese, French, and Japanese.

The open-weight release includes the H3-Base checkpoints, enabling local execution and fine-tuning. These weights provide the foundation for generating high-quality audiovisual content with 768p resolution on the shorter side. For optimal results, users should provide detailed descriptive prompts for both visual and auditory elements. When using reference assets, ensuring that the visual cues in images align with the motion described in the text prompt helps the model maintain character and scene consistency throughout the 15-second window.