Tencent releases HunyuanVideo
At 13 billion parameters, Tencent called it the largest open-weight video model, and said blind evaluators rated its motion quality above Runway Gen-3 and Luma 1.6.
- Open weights & ecosystem
- Models & capabilities
- Minor
Tencent released HunyuanVideo, a text-to-video foundation model published with open weights and inference code on GitHub, following the company’s earlier open release of its first Hunyuan large language model in 2023.
At roughly 13 billion parameters, Tencent described it as the largest openly released video-generation model at the time, built on a unified image-video transformer with a text encoder drawn from a multimodal large language model. The company said a professional evaluation, using more than 1,500 prompts judged by around 60 evaluators, scored HunyuanVideo above closed competitors including Runway’s Gen-3 and Luma’s 1.6 on motion quality and overall preference, though the comparison was Tencent’s own and not independently reproduced.
The release included a companion image-to-video variant and tools for avatar animation, and the company said architectural changes had cut training compute substantially against a naive scaling approach. Combined with the parameter count and permissive licence, it gave outside developers a video model competitive with the closed frontier — extending, in generative video, the pattern already visible in open-weight text models, where Chinese labs had begun matching or leading US releases in the tier just below the largest closed systems.