Alibaba’s Qwen family of open AI models is quietly taking the world by storm. Until the current month of January these models have overcome and to 700 million downloads on the Hugging Face platform. The milestone is significant and confirms the supremacy of Chinese companies in this type of models.
what has happened. Those responsible for the development of this family of models explain how Hugging Face data They don’t lie: these projects have become the most popular open models worldwide, at least if we look at the number of downloads.
Unstoppable. In October 2025 the Qwen model family managed to surpass the previous leader in this segment, the Llama de Meta family of models. Two months later, the Qwen models had been downloaded so much that the total number exceeded the combined figure of the eight AI models next in popularity. That group is made up of the models from Meta, DeepSeek, OpenAI, Mistral, NVIDIA, Zhipu.ai, Moonshot and Minimax.
Alibaba is a steamroller. Since Qwen launched in 2023, the advancement of the models in this family has been unstoppable. Although initially accessing them was more uncomfortableAlibaba has taken advantage of its infrastructure and its size to popularize them little by little, but above all its engineers have done something else: not stop launching models. The pace has been frenetic, but the models are also notable and comparable to proprietary models from major US technology companies.
Qwen wants to be the Android of AI. The family’s model catalog is enormous. Hugging Face currently includes 300 different models that cover both slightly older versions and various variants of each new major version. For example, we have models specialized in visual data recognition (Qwen VL), in programming (Qwen Coder) or models in image generation (Qwen Image Edit) or “generic” models absolutely gigantic or others like Omni that already compete with Grok 4 or GPT 5 Pro. The intention is obvious: to be the “standard” AI models on the market. Even if it is due to saturation.
Surprise: the most popular is not the most powerful. a study Published in October, it did provide surprising information on the growth in popularity of this family of models. One would expect that the most downloaded would be some of the versions of Qwen 3, the most modern and capable version. Actually the most downloaded is Qwen 2.5-1.5B-Instruct, a “light” model that can even run on modest mobile phones and laptops.
Tiny but bully. Currently the Hugging Face list indicates that the most popular in downloads is Qwen2.5-3B-Instruct, more modern and somewhat less lightweight, but still “small” by today’s standards. It seems clear that there is notable interest in being able to run this model on mobile phones, tablets and computers with little video memory.
Thousands of derived projects. The possibility of obtaining and using these open models in a simple way has made many developers and companies take advantage of them to customize them and adapt them to their own needs. That has made according to Xinhua That family has been used in more than 180,000 derived versions.
Flame fades. Meanwhile, Silicon Valley confirms that its vision is different. Goal, what led this area initially thanks to Llama, has given a rudder stroke. The company is still expected to launch new versions of this model, but in the meantime most technology giants focus on their closed and proprietary models. Of course: they keep some models open with a promising future, as we have seen with Gemma (Google) Phi (Microsoft) or gpt-oss (OpenAI). Without forgetting that Mistral is a great European benchmark and also offers open variants.
GIPHY App Key not set. Please check settings