MiniMax launches H3, promises open weights
MiniMax has released H3, a single model that takes text, images, video and audio as input and generates 2K video with sound, and says it will publish the weights in a field where closed models dominate.
MiniMax announced H3, a general-purpose multimodal generation model. It produces video clips of up to 15 seconds at 2K resolution with native stereo sound. Users can supply text, images, video and audio as context in one prompt. The company's example asks the model to borrow a "Hitchcock camera movement" from one video, animate a character from an image, and match the vocals to an audio clip.
MiniMax says it plans to release the weights "in the coming days, subject to applicable laws and regulations." It says the goals are to support the open-source community, speed up support for more kinds of AI hardware, and let users build customized versions. The company argues that closed-source models have slowed iteration in video generation.
On price, MiniMax says H3 costs less than a third of "mainstream models" per second at 2K. At 768p, it says H3 costs less than half what those models charge for 720p. It does not say which models it is comparing against.
H3 follows Hailuo 01 and Hailuo 02. Earlier models split work into separate tasks, such as text-to-video, image-to-video and reference-based generation, and treated voice, sound effects and music as separate domains. H3 trains on all of these together. Reference and editing instructions are written in natural language instead of being limited to a fixed set of tasks. The company lists Contextual Omni Representation, H3-VAE, an H3-Omni Transformer and In-Context Regeneration as its core technologies, and says better captioning was among its most important engineering work.
The post includes no benchmarks or independent comparisons. MiniMax has not said when the weights or a technical report will be released. The model is available to try through Hailuo AI.