MiniMax: hailuo-3

MiniMax H3 is an omni-modal generative model from MiniMax, released as open weights on August 3, 2026 under the MiniMax H3 Community License Agreement. It is built around the H3-Base core model, which combines a text encoder, a visual VAE, an audio VAE, and an omni transformer; RoPE is used to capture spatial and temporal relationships. The system also includes H3-Context-IR, a preprocessing and orchestration layer that is not part of the open-source release and is available only via API. A new high-compression tokenizer expands the effective sequence length fourfold and helps cut training and inference cost. The model accepts text, image, video, and audio inputs and understands the relationships between them. From a single model it generates video and audio together, at 480p, 768p, or native 2K resolution, at 24FPS and with lengths from 5 to 15 seconds. It supports first-frame and last-frame control, image-to-video conversion, video-to-video motion transfer, voice timbre reproduction from audio references, and native stereo audio covering voice, sound effects, and music. It follows complex multimodal instructions in natural language and handles editing tasks such as lip sync to new dialogue and partial reuse of background music, while rendering on-screen text and brand expression accurately. These abilities suit advertising, branding, e-commerce, product design, UI/UX, game content production, short-form video production with reference material, and instruction-based transitions. MiniMax describes the model as top-tier among open-source video generation models. Physical-world consistency remains a challenge, however: visual plausibility does not guarantee physical consistency, multimodal input does not always lead to effective multimodal reasoning, and reported failures include incorrect evidence grounding, incomplete event realization, and violations of contact mechanics and object geometry.