The Alibaba stand at the World Artificial Intelligence Conference at the Shanghai World Expo Exhibition Center in Shanghai, China, on July 5, 2024.
Nurphoto | Nurphoto | Getty Images
While U.S. markets have been primarily concerned with how Anthropic and Altruist’s tools impact software and financial services, Chinese tech giants have recently unveiled advanced AI models, highlighting notable progress in robotics and video generation.
Alibaba, ByteDance, makers of TikTok, and short-video platform Kuaishou, have all presented new AI models that demonstrate how Chinese companies are competing on a global scale.
This development follows comments from Google DeepMind CEO Demis Hassabis, who indicated in an interview with CNBC that Chinese AI models are only “months” behind their Western counterparts.
The newly released models from China are going head-to-head with video generation systems like OpenAI’s Sora and robotics initiatives by Nvidia and Google.
Here’s a summary of the latest models.
Alibaba’s RynnBrain
This week, Alibaba’s DAMO Academy introduced RynnBrain, an advanced AI model designed to enhance robot understanding of their physical surroundings and object recognition.
In a demonstration video, Alibaba showcased a robot equipped with pincered hands, capable of counting oranges, picking them up, and placing them into a basket. The robot was also observed retrieving milk from a fridge.
Creating such models demands substantial training to enable robots to recognize everyday objects, making even simple tasks like picking fruit quite challenging.
With RynnBrain, Alibaba now enters a competitive arena with companies like Nvidia and Google, both of which are developing their own robotics-focused AI models.
“One of its significant innovations is that it incorporates time and space awareness,” remarked Adina Yakefu, a researcher at Hugging Face, in an interview with CNBC.
“Rather than merely responding to immediate stimuli, the robot can remember when and where specific events occur, monitor task progress, and manage multiple steps. This makes it more dependable and coherent in complex real-world situations.”
Yakefu further emphasized Alibaba’s broader ambition to create a foundational intelligence layer for physical systems.
ByteDance’s Seedance 2.0
Seedance 2.0, introduced by ByteDance, is another video generation AI model capable of creating realistic videos from simple text prompts. These prompts can even include other videos and images.
Videos generated using Seedance 2.0, reviewed by CNBC, displayed strikingly realistic imagery and footage engineered entirely by AI.
Billy Boman, a creative advertising agency owner in Stockholm, Sweden, has worked with Seedance 2.0 and noted the substantial progress in AI video generation over the past two years.
“Back in 2023 … it was difficult to produce even a simple run or walk. Realistic output was often limited to very short clips, with poor textures and lack of detail. Now, everything has changed. I can now achieve impressive results in video production,” Boman remarked in a CNBC interview.
Hugging Face’s Yakefu pointed out that Seedance 2.0 has improved considerably over its predecessors in terms of “controllability, speed, and production efficiency.”
“In my experience, Seedance 2.0 is among the most versatile video generation models I’ve encountered. It truly surprised me with satisfactory results on the first attempt, even with a simple prompt. The visuals, music, and cinematography feel harmonious rather than experimental,” Yakefu stated.
Despite the positive feedback, Seedance has faced challenges. Reports from local Chinese media indicate that the feature allowing the AI to generate voices based on uploaded images has been suspended, following concerns raised by a blogger in China regarding the lack of consent for voice generation.
ByteDance did not respond to requests for comment from CNBC.
Kuaishou’s Kling 3.0
Kuaishou recently released Kling 3.0, a competitive video generation model aimed to rival ByteDance’s offerings.
The model features notable enhancements, including improved consistency, photorealistic output, extended video lengths of up to 15 seconds, and native audio generation in various languages, dialects, and accents.
Currently, Kling 3.0 is available solely to paying subscribers, but Kuaishou has plans to make it accessible to the public soon.
The success of Kuaishou’s Kling models has significantly contributed to a more than 50% increase in its share price over the past year.
Kuaishou shares year-to-date
Other key AI model releases
Zhipu AI, trading as Knowledge Atlas Technology on the Hong Kong exchange, experienced a surge in its shares last Thursday after revealing GLM-5, a cutting-edge open-source large-language model boasting advanced coding abilities and capabilities for extended agent tasks.
The company claims this model rivals Anthropic’s Claude Opus 4.5 in coding performance and outperforms Google’s Gemini 3 Pro in certain tests, although CNBC has yet to verify these assertions.
Shares of MiniMax also rose after it launched its updated M2.5 open-source model featuring enhanced AI agent tools, designed to automate various tasks.
—CNBC’s Anniek Bao and Dylan Butts contributed to this report.
