On a rainy afternoon this summer, I visited Rokid’s headquarters in Hangzhou, China, where the startup is creating innovative smart glasses. Through the use of a prototype device featuring cutting-edge technology, the engineers’ Mandarin was translated into English and displayed on a small screen positioned above my eye.
At the heart of this technology is Qwen (通义千问), an open-weight large language model developed by Alibaba, the leading Chinese e-commerce firm. While it may not boast the same performance metrics as alternatives like OpenAI’s GPT-5, Google’s Gemini 3, or Anthropic’s Claude, Qwen has gained traction, thanks to its ease of use and modifiability.
Recent data indicates that downloads of open Chinese AI models such as Qwen have surpassed those of US models on platforms like HuggingFace, reflecting a growing preference among developers. Models like DeepSeek have even released advanced language models requiring significantly less computational power than their US counterparts. By the end of the year, Qwen had positioned itself as the second-most popular open model globally.
Qwen’s capabilities, including product identification via camera, navigation, and web searching, align with user needs for advanced AI functionalities. Rokid has tailored Qwen to enhance the user experience within their smart glasses. Additionally, smaller versions of Qwen can be operated on smartphones, ensuring functionality even without internet connectivity.
During my visit, I had previously installed a compact version of Qwen on my MacBook Air, using it to aid in learning Mandarin. For many applications, such smaller open-source models deliver comparable performance to their larger, data center-based equivalents.
The growing success of Qwen contrasts with recent challenges faced by major American AI models. Meta’s Llama 4 release in April was criticized for underperformance relative to established benchmarks, prompting developers to explore alternative models. Similarly, the introduction of GPT-5 by OpenAI in August also disappointed, with users noting an uncharacteristically cold tone and several basic errors.
Despite OpenAI releasing a more streamlined open model, Qwen’s development pace and transparent engineering have contributed to its rising popularity among developers. Qwen has been featured in numerous academic papers at leading AI conferences, which highlights its prominence in the research community.
Not only are Chinese companies adopting Qwen for applications across various industries, but American firms are also integrating it into their systems. Companies like Airbnb and NVIDIA are reportedly leveraging Qwen, showcasing its extensive utility.
Experts suggest that the increasing focus on openness among Chinese AI firms, including extensive publication of research and engineering advances, offers a stark contrast to the more secretive nature of major US companies, which are more guarded about their technical processes.
The momentum of Qwen and similar models may also reflect a broader industry shift, emphasizing practical application and adaptability rather than merely excelling in narrow benchmarking measures. This trend indicates a growing recognition that real-world contributions of AI models are equally, if not more, important than their technical prowess in isolated scenarios, establishing Qwen and its counterparts as significant players in the evolving landscape of AI technology.
For more information on Qwen and its development: Alibaba | OpenAI | HuggingFace
