MANY MINDEDAn open model takes the top of a video ranking
MiniMax has released the weights for H3, its multimodal video model, The Decoder reported on 3 August 2026. Artificial Analysis, which runs independent model comparisons, ranks H3 first in Video Editing, second in Text-to-Video and third in Image-to-Video — the first time an openly available model has topped a video generation ranking.
H3 has 33 billion parameters and processes text, images, video and audio together, producing clips of four to fifteen seconds with stereo sound. Its model card describes prompts that can carry up to nine reference images, three video clips and three audio clips at once. The open weights permit fine-tuning on custom footage, characters or a particular visual style.
The reason this bears on abundance is the shape of the price. A capability behind an API costs whatever the vendor charges, indefinitely. The same capability as a downloadable file costs the hardware you run it on and then approaches zero per use — and it cannot be withdrawn, repriced or geofenced. That is the difference between renting a capability and owning one, and it is the mechanism by which frontier tools have historically reached people who could never have paid the rent.
The release is partial, and the limits deserve naming. Two pieces stay closed: the 2K resolution module, and H3-Context-IR, which converts prompts and reference material into a structured intermediate format. Running H3 locally in ComfyUI therefore tops out at 768p, and users must handle context preparation themselves using MiniMax's published prompting guides. The licence permits commercial use only for companies with under $20 million in revenue. ByteDance released its closed Seedance 2.5, which generates 30-second clips with built-in audio, the same day.
Source: The Decoder