Qwen

• 1. Introduction

  ›

  Over 200,000 derivative versions of Qwen exist on Hugging Face alone. A single vision model in the family, Qwen3-VL-2B-Instruct, had been downloaded more than 18 million times globally. Qwen is a family of large language models developed by Alibaba Cloud. It launched as a beta in April 2023 under the Chinese name Tongyi Qianwen. That September, regulatory clearance allowed it to open to the public.

  Those figures invite genuine questions. How did a company known primarily for e-commerce and cloud services build an AI model family with this global reach? What choices about licensing and architecture made Qwen so extensively forked by outside developers? And what do the leadership departures at Alibaba's AI division in 2026 reveal about the pressures shaping one of China's most visible AI projects?

  39828114

• 2. A Borrowed Blueprint

  ›

  Qwen's original model architecture was based on the Llama framework developed by Meta AI. In August 2023, the weights for the 7-billion-parameter version were released publicly. That December, Alibaba made two more sizes available for download: a 72-billion-parameter model and a 1.8-billion-parameter model.

  Qwen2 arrived in June 2024. It introduced both dense and sparse model variants to the family. September 2024 brought a partial opening: Alibaba released open weights for some Qwen2 models. Its most advanced Qwen2 variants remained proprietary, served only through Alibaba Cloud.

  On the 28th of April 2025, Alibaba released the Qwen3 family, with all models carrying the Apache 2.0 license. Its dense models came in six sizes: 0.6 billion, 1.7 billion, 4 billion, 8 billion, 14 billion, and 32 billion parameters. Its mixture-of-experts models reached 235 billion total parameters, with 22 billion activated per inference. The entire family was trained on 36 trillion tokens across 119 languages and dialects. Qwen3-Max, a proprietary model with over 1 trillion parameters and trained on approximately 36 trillion tokens, was accessible only through an API. Qwen3.5 and Qwen3.5-Plus followed on the 16th of February 2026, with Qwen3.5 released as an open-weight model.

  23343951

• 3. Open Weights, Forked Paths

  ›

  Alibaba released many Qwen models under the Apache 2.0 license. Other models carry the Qwen License or the non-commercial Qwen Research License. A subset of the most capable models are kept proprietary and served through Alibaba Cloud.

  By 2026, more than 100 open-weight Qwen models had been released. The entire collection had been downloaded more than 40 million times. That availability drew enthusiasts who built their own fine-tuned versions. One notable example is "Liberated Qwen," created by San Francisco-based Abacus AI. It is a version configured to respond to any user request without content restrictions. Qwen models are also available on ModelScope, a platform Alibaba built for open-source AI distribution.

  2830647882

• 4. Sensing the World

  ›

  The Qwen-VL series brought visual understanding to the Qwen family by combining a vision transformer with a large language model. Qwen2-VL was released in two sizes, 2 billion and 7 billion parameters. In January 2025, Qwen2.5-VL expanded that range to four variants: 3 billion, 7 billion, 32 billion, and 72 billion parameters. All except the 72-billion-parameter version were released under the Apache 2.0 license. Qwen-VL-Max, Alibaba's flagship vision model as of 2024, was priced at US$0.41 per million input tokens through Alibaba Cloud.

  On the 24th of March 2025, Qwen2.5-VL-32B-Instruct was released under the Apache 2.0 license as a successor. Two days later, on the 26th of March 2025, Qwen2.5-Omni-7B became available. It accepted text, images, video, and audio as inputs and could generate both text and audio as outputs, enabling real-time voice conversation. Distribution included chat.qwen.ai, Hugging Face, GitHub, and ModelScope.

  Qwen-Audio and Qwen2-Math extended the family to spoken language and numerical reasoning respectively. Qwen3-VL and Qwen3-Omni carry the multimodal approach into the Qwen3 generation.

  5378957

• 5. Machines That Reason

  ›

  QwQ-32B-Preview was released in November 2024 as a model focused on reasoning tasks, comparable in design to OpenAI's o1 system. On certain benchmarks, QwQ-32B-Preview outperformed o1. Its context window reached 32,000 tokens. Alibaba released it under the Apache 2.0 license. Only the model weights were made public; the training dataset and methodology were not disclosed.

  Qwen3 embedded reasoning into the model family's architecture through configurable modes. Qwen3 models could switch between a thinking mode and a non-thinking mode during use. Qwen3 Base served as the foundational pretrained model. Qwen3 Instruct supported only the non-thinking mode, while Qwen3 Thinking supported only the thinking mode. Qwen3-Max-Thinking, a proprietary reasoning variant, could generate text, pictures, and video.

  With Qwen3.5, Alibaba built a model designed to complete complex tasks. Alibaba claimed it could outperform U.S. rival models on objective metrics including speed and cost. It could independently take actions across mobile and desktop applications.

  4751151663

• 6. The Ecosystem Bet

  ›

  In November 2024, Alibaba launched the Accio application. Built on Qwen, Accio was designed for Alibaba's business-to-business e-commerce platform. It generated market insights and answered sourcing questions for the B2B marketplace. Tasks like data collection and trend tracking, previously done by hand, could be automated through the tool.

  A January 2026 update to the Qwen mobile application took that integration further. The update connected the Qwen chatbot directly to Alibaba's service ecosystem, beginning with food delivery. Alibaba announced plans to extend the integration to Taobao and Fliggy, allowing users to assign tasks to those platforms. The scope was to include consumer errands such as phone calls and document processing. Qwen's expanding role inside Alibaba's commercial infrastructure drew scrutiny from government analysts in the United States.

  492719397

• 7. A Contested Legacy

  ›

  China restricted access to Hugging Face in 2022. That decision led Alibaba to build ModelScope as an alternative distribution platform for open-source AI models. It gave Chinese developers another path to community-built models. Qwen's models are available on both platforms.

  The U.S.-China Economic and Security Review Commission assessed what this open-source approach had achieved. The commission found that China's development of open-source AI, including Qwen, had been critical to overcoming constraints in computing hardware. Expanded data-curation capacity, in the commission's view, was essential to China's ability to apply AI to research, manufacturing, and robotics.

  In March 2026, Lin Junyang, the head of the Qwen AI model division, resigned. He was the third senior Alibaba executive to leave that year. His departure followed the release of Qwen3.5 and Qwen3.5-Plus. Researchers expressed concern that the exits might signal a shift away from open-source AI development. The company stated publicly that its focus on open source would continue.

  That same month, Alibaba announced Alibaba Token Hub, a new AI business unit led by CEO Eddie Wu. It brought together the Tongyi Large Model Business Unit, the MaaS Business Line, Qwen, Wukong, and AI Innovation under one structure. Zhou Jingren became chief AI architect; CTO Wu Zeming joined the senior leadership. Tongyi Laboratory, the former name for the Tongyi Large Model Business Unit, was restructured to concentrate on Qwen development under Jingren's direction.

  In April 2026, Qwen3.5-Omni and Qwen3.6-Plus were released as proprietary products, with access limited to their chatbot interfaces and the Alibaba Cloud platform. That same month, Qwen3.6-35B-A3B was released under the Apache 2.0 license. The Qwen model family's published list already includes Qwen3-Next and Qwen3-Coder-Next, pointing to further releases in development.

  8272737576