By Roman Aglaganov
If you ask Western developers to name an open-weight model, many will say Llama. Ask what they are actually building on, and a surprising number will say Qwen.
From Tongyi Qianwen to Qwen
Alibaba Cloud launched its chatbot, Tongyi Qianwen, in 2023, and began releasing open-weight versions under the shorter name Qwen the same year. Unlike some rivals, it did not release one or two models. It released families: tiny models that run on a phone, mid-sized ones for a laptop, big ones for a data centre, plus versions specialised in code, maths and images.
Qwen3
In April 2025, Qwen3 arrived with eight open-weight models, from 0.6 billion to 235 billion parameters, all under the permissive Apache 2.0 licence, which allows commercial use. They were trained on about 36 trillion tokens covering 119 languages and dialects, and can switch between a “thinking” mode for hard problems and a fast mode for easy ones. The Qwen team claimed even its small 4-billion-parameter model could rival its previous 72-billion one.
The quiet winner
Qwen’s real success is as a foundation. There are more than 200,000 Qwen-based variants on Hugging Face, fine-tuned by companies, universities and hobbyists for their own languages and specialities. In November 2025, AI Singapore said it would switch its Southeast Asian language model, SEA-LION, from Llama to Qwen.
Why it matters
Open-weight AI used to mean American labs giving models away. Today, many of the most-used open models come from China, from Alibaba, DeepSeek, Moonshot and others. That gives developers worldwide more choice and cheaper tools, and it complicates the story of an AI race in which one country can simply out-build the other.
Worth knowing: models trained in China follow Chinese content rules, so they may be reticent on politically sensitive topics. For coding or customer support that rarely matters; for some uses it does.
