China’s push on Chinese AI models is set for a busy stretch as new releases and previews surface, including Alibaba’s next Qwen-3.5 family and what one report calls a strong “stealth” contender. The rapid pace is adding pressure in a market where companies are racing to stand out on performance, cost, and so-called agent-style capabilities that aim to carry out tasks across apps.
One report says the sector is “bracing for a monumental week,” with multiple new models appearing as domestic tech giants get ready to unveil flagship products. In that same report, Alibaba’s next-generation Qwen-3.5 is described as highly anticipated and timed for roughly a year after the prior Qwen-3 generation.
Alibaba’s Qwen-3.5 appears in previews
A South China Morning Post report says a member of Alibaba Cloud’s model-development team posted pull requests on Hugging Face and GitHub for a next-generation model family. The report describes Qwen-3.5 as the centerpiece of that family, and says the pull-request details point to two models—one with 9 billion parameters and another with 35 billion parameters—with native multimodal support for the first time in that line. It also says the models will use a next-generation architecture that was previewed in an experimental model called Qwen3-Next.
Another report claims larger specs and lower cost
A separate report says Alibaba unveiled Qwen3.5 on a Monday and positioned it for complex, autonomous tasks, while also claiming cost and performance improvements against major US competitors it listed as GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro. That report says Alibaba claimed Qwen3.5 runs 60% more cheaply than its predecessor and offers eight times higher throughput on large workloads, and it describes “visual agentic capabilities” that let the system execute actions across mobile and desktop applications. The same report says Alibaba published benchmark results showing Qwen3.5 outperforming earlier versions and the three Western systems “across several tests,” while also noting that in one benchmark category (GPQA Diamond) it scored 88.7 and placed third, and in another (IFBench) it scored 76.5 and led the compared systems.
That report also provides detailed technical claims, saying Qwen3.5 uses a hybrid architecture that combines linear attention via Gated Delta Networks with sparse mixture-of-experts routing. It describes Qwen3.5 as a 397-billion-parameter model that activates 17 billion parameters per forward pass, expands support to 201 languages and dialects (up from 119), and includes a Qwen3.5-Open-Source release with a 256,000-token context window, while keeping the largest Max-series models proprietary.
Conflicting descriptions on model size
The two reports describe Qwen-3.5’s scale differently. One says the new family includes two models at 9 billion and 35 billion parameters, based on preliminary information in pull requests, while the other describes Qwen3.5 as a 397-billion-parameter model with mixture-of-experts behavior.
Competition and limits in the AI race
The report on Qwen3.5 says Chinese technology firms are intensifying competition to attract users in a domestic market it describes as dominated by ByteDance’s Doubao chatbot and DeepSeek. It also says ByteDance released Doubao 2.0 on a Saturday, calling it an upgrade to a chatbot with nearly 200 million users in China, and adds that ByteDance framed its update around AI agent capabilities. The same report links the current burst of releases to what it calls the first anniversary of DeepSeek’s initial global breakthrough, which it says triggered a selloff in technology shares, and it says observers expect attention around a forthcoming DeepSeek announcement.
Even as Chinese developers push new models, a South China Morning Post report highlights concerns about whether China can surpass the US in the near term. It quotes Alibaba scientist Lin (as identified in the report) saying there is “less than 20%” chance China’s AI will exceed the US over the next three to five years, and it attributes part of the gap to US firms “pouring massive computational resources” into next-generation research while Chinese teams are stretched meeting daily demand. The same report says Tang Jie, co-founder and chief AI scientist at Zhipu AI (also known as Z.ai), supported the cautionary view on a panel at the AGI-Next summit hosted by Tsinghua University in Beijing’s Zhongguancun technology hub.
Still, that report says Chinese models have narrowed the performance gap with leading US models in recent years, citing third-party benchmarks. It also says US models have largely stayed closed-source while Chinese developers have “overwhelmingly” open-sourced their models, driving global adoption.
Wider stakes: diffusion, industry, and global links
A CSIS analysis frames the US and China as pursuing different “theories of value” in AI, with the US tied to frontier AI and large-scale investment, and China emphasizing targeted investments and competitive diffusion across the economy. The same analysis says China has a three-part structure focused on sector investment in core industries, reliable open-source platforms, and deeper AI integration across the economy, arguing that lower-cost Chinese models could help entry into developing economies and support future ecosystems and revenue.
CSIS also says survey-based comparisons of AI adoption in the US and China can be contradictory, while agreeing that adoption is accelerating. It points to areas where China’s deployment is likely strong, including industrial robotics—saying China had more than five times as many factory robots operating in 2024 (2,027,200 versus 393,700)—and notes China’s electricity generation scale and its large internet-user base. The same piece argues the US stands out in information technology such as cloud computing and cloud adoption, noting that Google, Microsoft, and Amazon together account for more than 50% of global cloud infrastructure market share, and it says China’s “relatively low rate of cloud adoption” complicates scaling upgrades and deployments.
Finally, CSIS says the US and China remain intertwined in finance and technology ecosystems, and it gives one example of cross-border reliance by stating that Airbnb uses Alibaba’s Qwen model for its customer service interface.
