The model race widened across weights, video and licensing
MiniMax, ByteDance and LG AI Research opened different distribution routes on the same day; competition now spans weight access, production environments and licence terms as well as capability.
Artificial Intelligence··Morning
Multimodality and a weights promise
MiniMax's announcement showed that the race is no longer moving along a single text-model axis. The company said H3, which handles text, images, video and audio together, can generate video at up to 2K resolution and 15 seconds in length with its own stereo audio. The announcement says the model can interpret context supplied in different formats jointly, edit existing content and transfer motion from one video to another. On pricing, the company said that at 2K its per-second price is less than a third of mainstream models, and at 768p less than half the price of mainstream models at 720p. MiniMax wrote that it plans to open the model weights in the coming days, subject to applicable laws and regulations, in order to support the open-source community and let the model run on a broader range of AI hardware. The announcement carries no comparative performance figures and no third-party evaluation results.[1]
The video model entered service directly
ByteDance widened the production format through video continuity on the same day. According to the company's announcement, the model can produce a 30-second high-quality video clip in a single run and supports multi-turn extension for longer sequences. The model accepts up to 30 images, 10 videos and 10 audio clips as reference material in a single input. Seedance 2.5 is being rolled out first to the Jimeng AI app and the Pro version of Doubao, with an API expected soon on Volcano Engine's Ark platform. TechNode's report cites the Chinese-language outlet Yicai as the source of the announcement. The report gives no figures for resolution or pricing and no comparison against rival models.[2]
Licensing became part of the release
LG AI Research presented scale together with licensing and downloadable weights. K-EXAONE 2.0 was released on Hugging Face under the Apache 2.0 licence, which permits commercial use. The company said the model has 750 billion parameters and that it completed the architecture, data training, distributed computing and inference infrastructure internally. The government-backed work aims to reduce dependence on foreign systems; company figures report stronger performance and wider language support, but no independent evaluation is yet supplied. These releases do not compete on one scoreboard. One stresses multimodality and a promise to open weights, another production video capabilities, and the third access to a large openly licensed model. A release is now defined not only by capabilities but by weight availability, production access and the licence boundary for reuse. Distribution also changes the stage at which the releases can be compared. MiniMax places open weights on its roadmap, Seedance is available directly as a service, and K-EXAONE supplies weights and commercial permission in the same release. A promise of access, a running service and a downloadable model are therefore different offers. Benchmark comparisons alone miss whether a model can run locally, enter a production chain or support derivative products.[3], [1], [2]
Related columns
For more information on this topic, you can read the related columns.
