What happened. On 31 July 2026 MiniMax announced H3, a general-purpose multimodal generation model that takes mixed text, image, video and audio context and produces video with native stereo sound, up to fifteen seconds long at 2K resolution. The announcement emphasised unified training across tasks that used to be separate models — text-to-video, image-to-video, editing, reference — and said the company planned to release the model weights in the following days, subject to applicable regulation.
Why it matters. H3 is the base model underneath the fastest public video endpoints in this category. Its open-weights release is what allowed fal to post-train and serve H3 Max, and H3 Max in turn is what today's continuous channels run on. Native audio is also notable for live formats: a channel that can generate sound with its picture is a different product from a silent one, although none of the channels we track has yet stated that its audio is model-generated.
Which platforms it affects. Indirectly, every platform running on H3 Max. Directly, none of our listings uses the base H3 endpoint, which is why the MiniMax H3 model brief on this site is marked as a base model rather than a deployed one.