Qwen (通义千问, Tongyi Qianwen) is Alibaba Cloud's family of large language, vision, audio and image models, built by Tongyi Lab. It is the most widely downloaded open-weight model family in the world and, since 2026, also the home of frontier-class hosted flagships. This site tracks the models, the releases, the numbers and the ways to actually use them.
August 2026 brought the biggest shift in Qwen's history: for the first time a Max-class model was released with downloadable weights, alongside a small dense model that beats the previous generation's hosted flagship.
Hosted flagship, 2 Aug 2026 - open weights, 12 Aug 2026
A 2.4-trillion-parameter mixture-of-experts model with roughly 95B parameters active per token.
Built for multi-day autonomous coding, research reproduction and long-horizon agent work, with a
reasoning_effort dial that trades depth against cost.
Released 14 Aug 2026
A dense 27B model under a clean Apache 2.0 licence that Qwen says outperforms the far larger hosted Qwen3.7-Plus on coding and office tasks. Handles text, images and video, with a thinking mode you can toggle per query the new default for single-GPU and local deployment.
Alibaba Cloud announced Tongyi Qianwen on 11 April 2023 as an enterprise chatbot for the Chinese market, and opened it to the public that September after regulatory approval. What started as one hosted assistant became a sprawling family: general chat models, dedicated coding models, reasoning models, vision-language models, end-to-end omni models that listen and speak, plus separate image and speech systems.
Two things make Qwen unusual. First, scale of openness Alibaba has published over a hundred open-weight checkpoints, most under Apache 2.0, and the ecosystem built on top of them now exceeds 200,000 derivative models on Hugging Face. Second, range the same generation ships models from under a billion parameters for phones up to trillion-parameter mixtures of experts for datacentres.
Since 2026 the strategy has been deliberately two-track: open weights for the community and for on-premise deployment, and proprietary Plus / Max tiers sold through Alibaba Cloud. The Qwen 3.8 release blurred that line by publishing Max-class weights for the first time.
General chat, long-document work and structured output across 119 languages, with explicit thinking modes for maths, science and multi-step logic.
Qwen3-Coder and the 3.x Max line target repository-scale work: tool use, terminal tasks, long-horizon autonomous coding and parallel agent orchestration.
The Qwen-VL line reads arbitrary-resolution images, documents, charts and long video, with spatial reasoning and visual grounding.
Qwen-Audio 3.0 covers recognition and generation hotword-aware ASR with strong domain-term recall, and expressive TTS in 16 languages.
Qwen-Image renders complex layouts and, unusually, accurate text inside images 12 languages, LaTeX pages and 100+ art styles in version 3.0.
Omni models take text, image, audio and video end-to-end and stream both text and speech back, enabling live voice conversation.
Launched in public beta on 17 November 2025, Alibaba's consumer app passed 100 million monthly active users within two months. Its distinguishing feature is agency rather than chat: it can order food through Taobao Instant Commerce, build and book travel through Fliggy, navigate with Amap, pay and access public services through Alipay, place phone calls and return transcripts, and batch-process up to 100 documents at a time.
"AI is evolving from intelligence to agency," as Alibaba Group VP Wu Jia framed the strategy.
Qwen is developed by Tongyi Lab inside Alibaba Cloud. In 2026 Alibaba consolidated its AI operations the Tongyi large-model unit, model-as-a-service and Qwen development into a group called Token Hub, reporting to CEO Eddie Wu.
The departure of Qwen division head Lin Junyang in March 2026, shortly after Qwen3.5, prompted questions about the durability of the open-source commitment. In practice Alibaba has kept publishing weights while monetising the top tier and Qwen3.8 went further than any previous release by opening Max-class weights.
Chat with it for free, call it from an API, or download the weights and run it yourself.