Podcast Episode
On Tuesday, Beijing-based Moonshot AI unveiled Kimi K2.5, an open-source multimodal model featuring a breakthrough "agent swarm" capability that can coordinate up to one hundred specialised sub-agents working in parallel. The Alibaba-backed startup claims the approach can complete complex tasks up to four and a half times faster than traditional single-agent systems. Built on a mixture-of-experts architecture with one trillion parameters and three hundred and eighty-four specialised experts, the model was trained on fifteen trillion mixed visual and text tokens.
Meanwhile, Baidu's shares surged to a near three-year high following the official release of Ernie five point zero, a model with two point four trillion parameters. The company's Ernie AI platform has now crossed two hundred million monthly active users.
The releases coincide with aggressive Lunar New Year marketing. Tencent announced it will distribute one billion yuan in cash prizes through its Yuanbao AI chatbot, while Baidu is offering five hundred million yuan in digital red envelopes through its Ernie Assistant app.
Demis Hassabis, CEO of Google DeepMind, recently told CNBC that Chinese AI models might be merely months behind those created in the United States.
Chinese AI Labs Unleash Wave of New Models, Closing Gap with US Rivals
January 28, 2026
Audio archived. Episodes older than 60 days are removed to save server storage. Story details remain below.
Chinese technology companies including Moonshot AI, Alibaba, and Baidu have released major new AI models this week, with capabilities that analysts say may be just months behind leading American systems. The releases coincide with aggressive Lunar New Year marketing campaigns offering hundreds of millions in cash prizes.
A Flood of New Models
Chinese artificial intelligence companies are making their most aggressive push yet to challenge American dominance in AI, releasing a wave of new models that suggest the technology gap between the two superpowers may be narrowing faster than expected.On Tuesday, Beijing-based Moonshot AI unveiled Kimi K2.5, an open-source multimodal model featuring a breakthrough "agent swarm" capability that can coordinate up to one hundred specialised sub-agents working in parallel. The Alibaba-backed startup claims the approach can complete complex tasks up to four and a half times faster than traditional single-agent systems. Built on a mixture-of-experts architecture with one trillion parameters and three hundred and eighty-four specialised experts, the model was trained on fifteen trillion mixed visual and text tokens.
Alibaba and Baidu Join the Race
The same day, Alibaba introduced Qwen3-Max-Thinking, its most advanced reasoning model to date. The system, trained on thirty-six trillion tokens with over one trillion parameters, can interleave tool calls within its reasoning process and features built-in web search capabilities. Alibaba says performance is comparable to Anthropic's Claude Opus four point five and Google DeepMind's Gemini three Pro across nineteen benchmarks.Meanwhile, Baidu's shares surged to a near three-year high following the official release of Ernie five point zero, a model with two point four trillion parameters. The company's Ernie AI platform has now crossed two hundred million monthly active users.
Open Source and Holiday Marketing
Chinese companies are pursuing distinctly different strategies than their American counterparts. Many models are open-sourced, offering free or low-cost access. According to recent reports, four out of five of the most-downloaded models on Hugging Face are now from Chinese labs.The releases coincide with aggressive Lunar New Year marketing. Tencent announced it will distribute one billion yuan in cash prizes through its Yuanbao AI chatbot, while Baidu is offering five hundred million yuan in digital red envelopes through its Ernie Assistant app.
Demis Hassabis, CEO of Google DeepMind, recently told CNBC that Chinese AI models might be merely months behind those created in the United States.
Published January 28, 2026 at 9:21pm
