GPT-5’s Out, Qwen’s In: The New AI Contender
▼ Summary
– The author visited Rokid, a Chinese startup, where engineers demonstrated smart glasses using a prototype that translated and transcribed Mandarin to English in real-time on a small screen.
– Rokid’s glasses are powered by Qwen, an open-weight AI model developed by Alibaba, which is not the top-performing model but is both capable and highly customizable.
– Chinese AI models like Qwen, DeepSeek, and others are gaining global popularity because they are effective and easy for developers to modify and use.
– Downloads of open Chinese AI models on HuggingFace surpassed those of US models in July, with Qwen becoming the world’s second-most-popular open model.
– The rise of Chinese models coincides with recent disappointments in US models like Meta’s Llama 4, leading developers to seek alternatives.
On a rainy, blustery summer day in Hangzhou, a visit to the smart glasses startup Rokid revealed a glimpse into the practical future of artificial intelligence. Through a prototype device, engineers’ Mandarin speech was instantly translated, transcribed, and displayed on a tiny screen before my eye. This seamless experience was powered not by a household-name American AI, but by Qwen, an open-weight large language model from Alibaba. While it may not top every benchmark against giants like GPT-5 or Claude, Qwen’s blend of high performance and remarkable accessibility is fueling a significant shift in the global AI landscape.
Qwen, known formally as Tōngyì Qiānwèn, is part of a wave of sophisticated Chinese models gaining substantial traction. They are not necessarily the absolute best in raw capability, but their open-weight nature makes them exceptionally versatile for developers and companies. Platforms like HuggingFace report that downloads for open Chinese models have now overtaken those for American ones. Another platform, OpenRouter, notes Qwen’s rapid ascent to become the world’s second-most-popular open model. This popularity stems from a powerful combination: these models are very capable and incredibly easy to adapt for specific needs.
For a company like Rokid, this adaptability is crucial. They can host a version of Qwen that’s been fine-tuned perfectly for their smart glasses, enabling features from visual product identification and map navigation to message drafting and web searches. The model’s flexibility even allows for compact versions to run directly on devices like smartphones, ensuring functionality persists without an internet connection. This practical utility extends to personal use as well; a smaller Qwen model installed on a laptop can serve as a capable assistant for tasks like language practice, proving that moderately sized open-source models can often match the performance of their massive, data-center-bound counterparts for everyday applications.
The growing appeal of Qwen and its peers, including DeepSeek and Moonshot AI, has been amplified by recent disappointments in the open-model space from other regions. When Meta released Llama 4 earlier this year, its performance failed to meet expectations on key benchmarks, leaving a void for developers seeking robust and malleable alternatives. This stumble coincided with Chinese labs demonstrating they could achieve cutting-edge results, as DeepSeek did, with reportedly far less computational power than their U.S. rivals. The result is a dynamic and increasingly competitive field where ease of use and customization are becoming just as important as raw benchmark scores.
(Source: Wired)
