Chinese AI companies are using the Lunar New Year period to roll out new models and features, aiming for another breakout moment like DeepSeek’s earlier impact. The releases span chatbots, “agent” tools that can carry out tasks, and fast-moving upgrades in video and image generation.
One of the biggest talking points has been GLM-5, which was first spotted as a mysterious test model called Pony Alpha before being revealed as a product from Beijing-based Zhipu AI. A post referenced by Xinhua said the model performed strongly on agent benchmarks and was used in “many agent flows” during its stealth-testing period.
GLM-5 steps into the spotlight
Xinhua reported that Pony Alpha was revealed to be GLM-5 by Zhipu AI after days of speculation in the AI community. It also described stealth testing as a sign of growing confidence among Chinese developers in their products.
In a separate write-up, Latent Space said GLM-5 is positioned for long-horizon, agent-focused work and outlined technical changes it attributed to a GLM-5 blog post. Latent Space said GLM-5 scales from 355B parameters (32B active) to 744B parameters (40B active) and increases pre-training data from 23T to 28.5T tokens. The same report said GLM-5 integrates DeepSeek Sparse Attention (DSA) to reduce deployment costs while keeping long-context capability.
Latent Space also said that discussion around the release included constraints on serving capacity, describing “compute” as tight and warning that inference capacity was limited. It also cited figures shared in posts that mentioned a 200K context limit and 128K maximum output.
Kimi 2.5 and DeepSeek upgrades raise the pressure
Xinhua said Moonshot, another Beijing tech firm, unveiled a new Kimi 2.5 release late last month that surprised the AI community with open-source performance across agents, coding, image, video, and general intelligence tasks. It also quoted Chamath Palihapitiya calling the open-source model “incredibly profound” while saying it would save 90 percent of the cost and create a “Kimi 2.5 moment.”
DeepSeek, meanwhile, has continued testing changes to its own systems. Xinhua reported that DeepSeek started a gray-scale test on its app and web interfaces and expanded its context window length from 128K tokens to one million tokens.
Xinhua framed the broader wave of launches as an effort to replicate the global impact of DeepSeek’s R1 launch in January 2025. It said that push is also spreading beyond text systems into video and image AI.
Video and image tools join the New Year sprint
On the video side, Xinhua said ByteDance tested a text-to-video tool called Seedance 2.0 that can generate a multi-shot film sequence in roughly 60 seconds using relatively simple prompts. Feng Ji, identified by Xinhua as the creator of the game Black Myth: Wukong, was quoted saying, “The childhood era of AIGC is over,” while discussing Seedance 2.0. Xinhua also quoted him adding, “It’s fortunate that, at least for now, Seedance 2.0 originates from China.”
On the image side, Xinhua reported that Alibaba’s AI arm launched its next-generation image generation foundation model, Qwen-Image 2.0, on Tuesday. It also said the open-source Qwen model is powering AI tools for the International Olympic Committee as it prepares for the 2026 Milan Winter Olympics.
Baidu brings an AI agent into search
The competition is also showing up inside everyday consumer apps. CryptoRank reported that Baidu integrated an AI agent called OpenClaw into its main smartphone search app, giving it access to a large user base that the article described as about 700 million monthly active users. The same report said the integration is designed to help users do tasks such as scheduling, organizing files, and writing code without needing a separate chat app.
CryptoRank said Baidu planned to let users who opt in message OpenClaw directly through the search app starting later on Friday. It described OpenClaw as an Austrian-developed, open-source AI agent that previously was available through messaging platforms such as WhatsApp and Telegram.
Alongside the rollout, CryptoRank reported that cybersecurity concerns are being raised as AI agents get deeper access to systems and data. It said CrowdStrike warned users about granting unrestricted access to enterprise systems.
Alibaba’s shopping chatbot push and China’s user numbers
CryptoRank also said Alibaba integrated its AI chatbot Qwen into consumer services including Taobao and its travel service Fliggy. The report said Alibaba claimed the chatbot helped process more than 120 million consumer orders during the six days through Feb. 11, and that Qwen lets users compare personalized recommendations and complete purchases directly through Alipay within the chatbot. It added that users previously had to leave the chatbot and switch platforms to finish a transaction.
Xinhua linked these product pushes to a wider effort by major firms to lock in users for intelligent assistants during the holiday season. It said companies including Alibaba, Tencent, and ByteDance are working to accelerate what it called an “R&D-to-consumption ecosystem” built around their AI models and platforms.
Xinhua also cited figures from the China Internet Network Information Center showing China had 1.125 billion internet users by the end of 2025. It said users of generative AI technology reached 602 million, a 141.7 percent increase from the end of 2024.
