Vibeleaderboard
Qwen

ai lab

Also indexed as alibaba-qwen · qwenlm

Qwen

Qwen matters because it connects a broad, frequently updated open-weight model portfolio to Alibaba's cloud and product distribution. It gives researchers and deployers reusable bases across language, code, vision, audio, and retrieval, making it one of the clearest examples of a large technology company treating open model releases as an ecosystem strategy rather than a one-time publication.2,4,5,1

Official site ↗Models in the Atlas →

Profile

Overview

From Tongyi Qianwen to Qwen

Qwen is Alibaba's foundation-model program, first introduced in April 2023 as Tongyi Qianwen. Alibaba Cloud initially positioned it as a general model for DingTalk, Tmall Genie, and customer applications built through its cloud platform. The English-language identity later consolidated around Qwen, while Tongyi remains part of the program's Chinese branding and product history.1,8

A portfolio rather than one model

The program evolved from a single language-model line into a portfolio spanning dense and mixture-of-experts language models, code, vision-language, audio, speech recognition, image generation, embeddings, and reranking. Qwen2 and Qwen2.5 reports documented multilingual and long-context improvements across multiple parameter sizes. Qwen3 added a hybrid design that can switch between explicit thinking and faster non-thinking modes.2,3,4

Open weights and cloud distribution

Open weights are central to Qwen's reach. Alibaba distributes models through GitHub, Hugging Face, ModelScope, and Alibaba Cloud, allowing local deployment, fine-tuning, and hosted use. The licenses are model-specific, so the portfolio should not be described as uniformly open source. Its practical ecosystem influence comes from the number of sizes and modalities available to downstream builders, not from a claim that Qwen invented multimodal or open-weight modeling.5,6,2

Leadership and licensing transition

The organization entered a leadership transition when Qwen division head Junyang Lin resigned in March 2026. Reuters reported that Alibaba created a group-wide task force coordinated by chief executive Eddie Wu, group technology chief Wu Zeming, and Alibaba Cloud technology chief Jingren Zhou. Reuters later reported planned revenue-sharing terms for major commercial users of a future Qwen model. The first event is confirmed organization news; the precise long-term license model remains uncertain because the reported plan was not yet public policy.9,10

Reporting and context

Notable contributions

  1. 01A coordinated open-weight model portfolioQwen developed reusable model families across language, code, vision, audio, image generation, embeddings, and reranking, with multiple sizes and deployment routes. The contribution is breadth and coordination, not priority over the underlying modalities.5,6
  2. 02Hybrid thinking controls in Qwen3Qwen3 exposed thinking and non-thinking behavior within one model family so applications could trade additional reasoning for latency and cost according to the task.4
  3. 03Multilingual open models across sizesQwen2 and Qwen2.5 released a ladder of models with expanded multilingual coverage and context lengths, giving local and commercial deployers more choices than a single frontier-size artifact.2,3
  4. 04Cloud and local distribution from one programAlibaba paired downloadable Qwen weights with hosted models and enterprise tooling in Alibaba Cloud Model Studio, supporting both customer-controlled and managed deployment.1,3