Qwen

Qwen (Tongyi Qianwen, Chinese: 通义千问, “Truth from a Thousand Questions”) is a public brand of AI models developed by the Qwen Team within Alibaba Cloud, a division of the Chinese technology corporation Alibaba Group Holding Limited, founded on June 28, 1999, in Hangzhou.

The history of Qwen began in April 2023 with the announcement of its first model, Tongyi Qianwen, at the Alibaba Cloud Summit in Beijing. Following this release, the company rapidly launched over a hundred open and commercial models of varying sizes — from compact 0.5B-parameter versions for mobile devices to massive MoE models exceeding one trillion parameters. However, Qwen became renowned not only for its productivity but also for setting industry trends and standards, consistently introducing innovative solutions that have become benchmarks in their respective categories — from programming and mathematics to image, video, and audio analysis.

Moreover, Qwen demonstrates remarkable agility for a company of its scale, actively experimenting and swiftly adapting its products based on community feedback. A striking example of this is the evolutionary path of its flagship Qwen3 series, introduced in April 2025. Initially, it presented a novel hybrid reasoning concept with switchable “thinking” and “non-thinking” modes within a single model, allowing users to dynamically control the depth of reasoning via an API parameter. However, responding to developers’ practical needs and aiming to simplify deployment, the Qwen team rethought this approach within just a few months after the release: the “Thinking” and “Instruct” modes were formalized and separated into distinct, clearly specialized models within the Qwen3 series.

Qwen’s key products and research form a comprehensive ecosystem, where each release introduces fundamentally new technological breakthroughs. The Qwen3 series pioneered the aforementioned hybrid reasoning concept. QwQ, launched in November 2024, became the first open-source medium-sized model capable of successfully competing with closed systems in complex logical and mathematical tasks. In the field of multimodality, the QvQ-72B model, released in December 2024, became the first open-weight visual reasoning model capable not only of recognizing objects but also of generating detailed textual explanations of its own “thought process.” This high-quality image understanding was further advanced by the Qwen2.5-VL series, which added support for videos up to one hour long with precise temporal localization of events. Another significant achievement is Qwen2.5-1M, one of the world’s first open models supporting a context window of 1 million tokens. Additionally noteworthy are the Qwen-Image and Qwen-Image-Edit services, which enable not only image generation but also precise, detail-preserving image editing.

Overall, Qwen is one of the main drivers of the open-source AI industry, as the vast majority of its models are open-weight and distributed under the permissive Apache 2.0 license, allowing unrestricted commercial use. On the Hugging Face platform, more than 20 model collections are available, including specialized models for programming, mathematics, and multimodal tasks involving images, video, audio, and more. This exceptional openness, combined with enterprise-grade model quality, comprehensive and clear documentation, and active community support, has made Qwen one of the most popular, sought-after, and respected ecosystems in the world.

Related models

QwQ