It’s been a big — but rocky — week for AI models from China. Here’s what’s happened

While U.S. markets have been focused on the impact of Anthropic and Altruist’s tools on software and financial services, China’s tech giants have released AI models this week that have shown advancements in robotics and video generation.

Alibaba, TikTok creator ByteDance and short-video platform Kuaishou, have all released new AI models that underscore how Chinese firms are keeping up with those in the U.S.

It comes after Google DeepMind boss Demis Hassabis told CNBC that Chinese AI models are just “months” behind Western rivals.

These models from China are directly competing with video generation models such as OpenAI’s Sora, as well as robotics models from Nvidia and Google.

Here’s a rundown of the models.

Alibaba’s DAMO Academy unveiled RynnBrain this week, an AI model designed to help robots comprehend the physical world around them and identify objects.

In a video demo, Alibaba showed a robot with pincers for hands that appeared to be able to count oranges, pick them up and place them in a basket. It was also shown taking milk out of a fridge.

Models require extensive training to enable them to identify everyday objects to interact with, which means that simple tasks like picking up fruit can be challenging in robotics.

RynnBrain now puts Alibaba in competition with the likes of Nvidia and Google which are developing their own AI models for robots.

“One of its key innovations is built-in time and space awareness,” Adina Yakefu, a researcher at Hugging Face, told CNBC.

“Instead of simply reacting to immediate inputs, the robot can remember when and where events occurred, track task progress, and continue across multiple steps. This makes it more reliable and coherent in complex real-world environments.”

Yakefu added that Alibaba’s “broader ambition” was to “establish a foundational intelligence layer for embodied systems.”

Seedance 2.0 is a video generation AI model capable of generating a realistic video from just a text prompt from a user. But prompts can also contain other videos and images.

Videos created with Seedance 2.0 and reviewed by CNBC appear to show quite realistic imagery and video that has been fully created with AI.

Billy Boman, who is based in Stockholm, Sweden, and runs a creative advertising agency that produces AI-generated content, has used Seedance 2.0.

He said AI video generation has made significant strides over the past two years, with rapid improvements across the industry.

“Back in 2023 … it was difficult to get someone to run or to walk. Any type of realism was [limited to] very short clips, everything was very slow, bad textures, no skin textures, lacking detail. Now the script has flipped. Now I can do anything. It has been nothing short of exceptional, the technological advancements,” Boman told CNBC in an interview.

Hugging Face’s Yakefu, added that the Seedance 2.0 model has shown progress from previous generations in “controllability, speed and production efficiency.”

“Seedance 2.0 is one of the most well-rounded video generation models I’ve tested so far. It genuinely surprised me by delivering satisfying results on the first try, even with a simple prompt. The visuals, music, and cinematography come together in a way that feels polished rather than experimental,” Yakefu said.

However, while users have praised the technology, Seedance has run into trouble. Local Chinese media reported that Seedance has suspended a feature that allowed the AI to generate the voice of a person based on a picture they uploaded. It came after a blogger in China raised concerns about the voice generation taking place without consent.

ByteDance was not immediately available for comment when contacted by CNBC.

Released last week, Kuaishou’s Kling 3.0 is another video generation model to rival ByteDance’s.

Kling 3.0 “features major upgrades in consistency, photorealistic output, extended video duration up to 15s, and native audio generation across multiple languages, dialects, and accents.

The model is only available to paying subscribers but will be available to the public soon, Kuaishou said.

Kuaishou’s success with its Kling models has been a key factor behind its more than 50% share price rise over the last year.

Zhipu AI — which trades as Knowledge Atlas Technology in Hong Kong — saw its shares surge on Thursday after it released GLM-5, an open-source large-language model with enhanced coding capabilities and long-running agent tasks.

The company said the model approaches Anthropic’s Claude Opus 4.5 in coding benchmarks while surpassing Google’s Gemini 3 Pro on some tests. CNBC could not verify those claims.

Shares of MiniMax also jumped Thursday after it launched its updated M2.5 open-source model with enhanced AI agent tools. “Agents” or “agentic AI” refers to AI tools designed to automate tasks.

— CNBC’s Anniek Bao and Dylan Butts contributed to this report.