Tencent Heats Up Ai Video-generation Competition In China With New Open-source Product

Competition is heating up in the field of artificial intelligence (AI) video-generation technology in China, after internet giant Tencent Holdings launched its new open-source HunyuanVideo-I2V model to developers.

The image-to-video model, based on Tencent’s open-source HunyuanVideo foundation model introduced in December, lets users turn a static photo into a 720-pixel, high-resolution video clip of up to 129 frames, or five seconds, via short text prompts.

Shenzhen-based Tencent, which runs the world’s largest video-gaming business by revenue and China’s biggest social-media network, made the new model widely available on Thursday through online developer platforms like GitHub and HuggingFace.

Users can also add lip-synched voice to their videos as well as background sound effects, according to Tencent. In a demonstration video released by the company, Albert Einstein was shown eating an apple with a crunching sound effect.

China is seeing its field of home-grown AI video-generation products get crowded, more than a year after ChatGPT creator OpenAI astonished developers around the world with its Sora text-to-video model.

That field includes highly competitive products from Kuaishou Technology, TikTok owner ByteDance and Alibaba Group Holding’s cloud computing services unit. Alibaba owns the Post.

When HunyuanVideo was made available for free to businesses and individual users in December, Tencent claimed that it was the world’s largest open-source model for video generation, with over 13 billion parameters – a measure of variables present in an AI system during training.

The landing page of Tencent Holdings’ HunyuanVideo foundation model, which was used to develop the company’s new open-source HunyuanVideo-I2V model. Photo: SCMP

Alibaba Cloud, the cloud computing and AI arm of Alibaba, last month announced that it would open source four models of its Wan2.1 series – the latest version of its video-generation foundational model Tongyi Wanxiang – making them freely available to academics, researchers and commercial institutions worldwide.

Released in January, the Wan2.1 series was touted as the first video-generation model to support text effects in both Chinese and English. As of Monday, the Wan 2.1 model ranked as the top video-generation model on the VBench Leaderboard, a benchmark test for video generation, and was the only open-source model among the top five.

ByteDance last month unveiled its OmniHuman-1 multimodal model, which went viral for its ability to transform photos and sound bites into realistic videos. In September, the company launched two video-generation AI models, Doubao-PixelDance and Doubao-Seaweed.

Short-video platform operator Kuaishou, the main rival of ByteDance-owned Douyin in China, in June launched its Sora-like Kling AI Model. It was later upgraded with new iterations and offered for subscription. As of January, Kling had over 6 million users who generated more than 65 million videos and 175 million images, according to the company.

Beijing is also looking to accelerate development in video-generation technology amid the growing AI rivalry between China and the US.

At a press conference about the latest central government work plan on Wednesday, researcher Chen Changsheng – who took part in drafting that document – said Kuaishou’s Kling AI model had already surpassed the capabilities of Sora to some extent based on feedback from the global developer community.

The Hong Kong-listed shares of Kuaishou rose nearly 5 per cent on Friday after surging 15 per cent on Thursday. Tencent’s shares closed down about 2 per cent on Friday.

Meanwhile, the open-source approach to AI model development has gained momentum on the mainland, thanks in part to Hangzhou-based DeepSeek’s recent breakthrough. The start-up built its advanced open-source AI models, V3 and R1, at a fraction of the cost and computing power that major tech companies typically require for large language model (LLM) projects.

LLMs are the technology underpinning generative AI services like ChatGPT and DeepSeek’s namesake chatbot. An open-source project gives public access to a program’s source code, allowing third-party software developers to modify or share its design, fix broken links or scale up its capabilities.