China's AI model scale race

ByteDance developing AI model with up to 10 trillion parameters

ByteDance released Seedance 2.5, a video model capable of generating 30-second videos with multimodal inputs.

On July 31, 2026, ByteDance released Seedance 2.5, a video model capable of generating videos up to 30 seconds in length. The model accepts various multimodal inputs, including image, video, audio, and text.

This release followed the earlier release of Seedance 2.0 in early 2026, which generated realistic AI videos that attracted attention and raised copyright concerns from Hollywood studios.

ByteDance released Seed2.1, its top general-purpose multimodal model, designed for reasoning and agentive tasks.

No separate article for this desk or language yet.

Meituan and DeepSeek released models with approximately 1.6 trillion parameters each, among China's largest at the time.

Prior to the development of newer models, Meituan's LongCat-2.0 and DeepSeek V4-Pro were recognized as some of the largest models in China. Each of these models was reported to have approximately 1.6 trillion parameters.

Moonshot AI released Kimi K3, a model with 2.8 trillion parameters, setting a new benchmark in China.

Moonshot AI developed the Kimi K3 model, which contains 2.8 trillion parameters.

At the time of its release, the model represented a significant scale in the Chinese AI landscape.

ByteDance released Seedance 2.0, a video generation model that drew copyright concerns from Hollywood studios.

In early 2026, ByteDance released Seedance 2.0, a video generation model capable of producing realistic AI videos. The release drew significant attention and raised copyright concerns among Hollywood studios.