Farewell, GPT-5. Welcome, Qwen!

On a gray and breezy summer afternoon, I had the opportunity to visit Rokid’s headquarters, a startup innovating in smart glasses in Hangzhou, China. During my discussions with the engineers, their commentary was rapidly translated from Mandarin to English, then displayed on a small translucent screen positioned just above my right eye via one of the company’s latest prototype devices.

Rokid’s advanced eyewear utilizes Qwen, an open-weight large language model created by the Chinese e-commerce titan Alibaba.

Qwen—known in Chinese as 通义千问 or Tōngyì Qiānwèn—is not considered the top AI model available. Models such as OpenAI’s GPT-5, Google’s Gemini 3, and Anthropic’s Claude typically achieve higher scores on various benchmarks measuring machine intelligence. Additionally, Qwen isn’t the first leading-edge open-weight model; that distinction goes to Meta’s Llama, which debuted in 2023.

However, Qwen and other Chinese models—from DeepSeek, Moonshot AI, Z.ai, and MiniMax—are gaining traction due to their impressive capabilities and user-friendly nature for modifications. According to HuggingFace, a platform offering access to AI models and code, Chinese open models outpaced U.S. downloads in July this year. DeepSeek made waves by unveiling a state-of-the-art large language model that operates with significantly less computing power than its U.S. counterparts, while OpenRouter reports that Qwen has surged in popularity over the past year, becoming the second-most-popular open model globally.

Qwen is capable of handling most requests one would expect from an advanced AI model. For Rokid’s users, this may encompass tasks such as recognizing products captured by an integrated camera, providing directions via a map, composing messages, or conducting web searches. Given that Qwen can be effortlessly downloaded and adjusted, Rokid offers a tailored version of the model optimized for its applications. There’s also the capability to run a small-scale version of Qwen on smartphones or other devices to ensure functionality even when internet access is interrupted.

Prior to my trip to China, I installed a compact version of Qwen on my MacBook Air to practice basic Mandarin. For numerous applications, smaller open-source models like Qwen perform comparably to the larger models housed in vast data centers.

The emergence of Qwen and other Chinese open-weight models parallels the setbacks experienced by some prominent American AI models in the past year. When Meta introduced Llama 4 in April 2025, the model’s performance disappointed many, falling short of performance standards set by popular benchmarks like LM Arena. This disappointment has prompted many developers to seek alternative open models for experimentation.