So Long, GPT-5. Hello, Qwen

On a drizzly and windswept day this summer, I visited the office of Rokid, a startup processing astute glasses successful Hangzhou, China. As I chatted with engineers, their words were swiftly translated from Mandarin to English, and past transcribed onto a tiny translucent surface conscionable supra my close oculus utilizing 1 of the company’s caller prototype devices.

Rokid’s high-tech spectacles usage Qwen, an open-weight ample connection exemplary developed by the Chinese ecommerce elephantine Alibaba.

Qwen—full sanction 通义千问 oregon Tōngyì Qiānwèn successful Chinese—is not the champion AI exemplary around. OpenAI’s GPT-5, Google’s Gemini 3, and Anthropic’s Claude often people higher connected benchmarks designed to gauge antithetic dimensions of instrumentality cleverness. Nor is Qwen the archetypal genuinely cutting-edge open-weight model, that being Meta’s Llama, which was released by the societal media elephantine successful 2023.

Yet Qwen, and different Chinese models—from DeepSeek, Moonshot AI, Z.ai, and MiniMax—are progressively fashionable due to the fact that they are some precise bully and precise casual to tinker with. According to HuggingFace, a institution that provides entree to AI models and code, downloads of unfastened Chinese models connected its level surpassed downloads for US ones successful July of this year. DeepSeek shook the satellite by releasing a cutting-edge ample connection exemplary with overmuch little compute than US rivals, but OpenRouter, a level that routes queries to antithetic AI models, says Qwen has rapidly risen successful popularity done the twelvemonth to go the second-most-popular unfastened exemplary successful the world.

Qwen tin bash astir things you’d privation from an precocious AI model. For Rokid’s users, this mightiness see identifying products snapped by a built-in camera, getting directions from a map, drafting messages, searching the web, and truthful on. Since Qwen tin easy beryllium downloaded and modified, Rokid hosts a mentation of the model, fine-tuned to suit its purposes. It is besides imaginable to tally a teensy mentation of Qwen connected smartphones oregon different devices conscionable successful lawsuit the net transportation goes down.

Before going to China I installed a tiny mentation of Qwen connected my MacBook Air and utilized it to signifier immoderate basal Mandarin. For galore purposes, modestly sized unfastened root models similar Qwen are conscionable arsenic bully arsenic the behemoths that unrecorded wrong large information centers.

The emergence of Qwen and different Chinese open-weight models has coincided with stumbles for immoderate celebrated American AI models successful the past 12 months. When Meta unveiled Llama 4 successful April 2025, the model’s show was a disappointment, failing to scope the heights of fashionable benchmarks similar LM Arena. The gaffe near galore developers looking for different unfastened models to play with.