MiniMax Releases MiniMax H3: An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio
MiniMax has released H3, a model the company describes as a general-purpose multimodal generation system. Unlike conventional text-to-video tools that treat audio as an afterthought or a separate post-processing step, H3 is designed from the ground up to ingest and output multiple modalities together - text, images, video, and audio are all handled within one unified context window, and the resulting video clip includes native stereo sound rather than mono or silence.
On the output side, H3 generates clips at 2K resolution with durations ranging from 4 to 15 seconds in integer steps. That ceiling of 15 seconds is notably longer than what many competing models currently offer at comparable resolution, and the stereo audio output is a meaningful differentiator at a time when most video generation models still produce silent footage that requires a separate audio pipeline to be useful in production.
The architecture framing - treating all modalities as one context rather than chaining specialized sub-models - has practical implications for consistency. When a model processes a reference audio clip, a style image, and a text prompt together, the resulting video is less likely to show the seams that appear when separate systems are stitched together after the fact. This approach mirrors design decisions seen in some large language models that handle mixed inputs, now applied to a generation rather than comprehension task.
MiniMax has been building out its video generation capabilities steadily, having previously released models in its Video-01 line. H3 represents a more ambitious architectural step, moving away from a video-first model with optional extras toward something closer to a native omni-modal system. Details on the underlying architecture, training data, and API availability were not fully disclosed at launch, but the model appears aimed at developers and creators who need a single endpoint capable of handling complex, mixed-modality prompts without assembling multiple tools.
