Qwen, also known as Tongyi Qianwen (Chinese: 通义千问; pinyin: Tōngyì Qiānwèn), is a family of predominantly open-weights large and small language models (LLMs and SLMs) developed by Alibaba Cloud. The models are designed for a range of natural language processing tasks, including text generation, reasoning, and multimodal understanding. Qwen has become a significant player in the global Artificial intelligence landscape, particularly noted for its permissive licensing and the wide adoption of its smaller models within the open-source community.
The family has expanded through several major iterations, from the initial Qwen release in 2023 to the advanced Qwen3.8 series in 2026. These models are built on Transformer (architecture) architectures and leverage techniques from Deep learning and Machine learning. Qwen models are frequently used as a starting point for further fine-tuning and customization, with many derivative models created by independent developers.
Origins and Initial Release
Alibaba launched a beta of Qwen in April 2023 under the name Tongyi Qianwen. The initial architecture was based on Meta AI's Llama 1 model, a foundational Large language model design. After receiving regulatory clearance, Alibaba opened Qwen for public use in September 2023.
The first open-weights releases occurred in August 2023 with the 7B parameter model, followed by the 72B and 1.8B variants in December 2023. These early releases established Qwen's presence in the open-source AI community, providing researchers and developers with accessible models for experimentation.
Qwen2 Series
Qwen2 was released in June 2024, introducing both dense and sparse (mixture-of-experts) model architectures. In September 2024, Alibaba released some Qwen2 models with open weights while keeping its most advanced models proprietary. The series included specialized variants such as Qwen-VL for visual language tasks and Qwen-Audio for audio processing.
The Qwen2-VL line combined a vision transformer with an LLM, with variants of 2 billion and 7 billion parameters. In January 2025, Qwen2.5-VL expanded this with 3, 7, 32, and 72 billion parameter versions, all licensed under the Apache 2.0 license except the 72B variant. Qwen-VL-Max, Alibaba's flagship vision model as of 2024, was sold through Alibaba Cloud at US$0.41 per million input tokens.
In November 2024, Alibaba released QwQ-32B-Preview, a reasoning-focused model similar to OpenAI's o1, under the Apache 2.0 License. This model had a 32K token context length and performed better than o1 on some benchmarks. The Accio application was also launched that month.
On 29 January 2025, Alibaba launched Qwen2.5-Max. On 24 March 2025, Qwen2.5-VL-32B-Instruct was released as a successor to the earlier vision model. Two days later, Qwen2.5-Omni-7B was released under the Apache 2.0 license, accepting text, images, videos, and audio as input while generating both text and audio output, enabling real-time voice chatting.
Qwen3 Family
The Qwen3 model family was released on 28 April 2025, with all models licensed under the Apache 2.0 license. This release included both dense and mixture-of-experts (MoE) models. Dense model sizes ranged from 0.6B to 32B parameters, while MoE models included 30B-A3B (30B total with 3B activated) and 235B-A22B (235B total with 22B activated). These models were trained on 36 trillion tokens across 119 languages and dialects.
The Qwen3 collection included several variants: Qwen3 with switchable thinking and non-thinking modes, Qwen3 Base for pretraining, Qwen3 Instruct for non-thinking mode only, and Qwen3 Thinking for thinking mode only. Additionally, Qwen3-Max, a proprietary model with over 1 trillion parameters, was available through API, along with Qwen3-Max-Thinking, a reasoning variant capable of generating text, pictures, or video.
Qwen3.5 and Later Releases
In February 2026, Alibaba released the open-weights Qwen3.5 and the proprietary Qwen3.5-Plus. Alibaba stated that Qwen3.5 could operate desktop and mobile applications. Qwen3.5-Omni and Qwen3.6-Plus were released in April 2026 as proprietary models, with access limited to chatbot websites and the Alibaba cloud platform. The Qwen3.6 model was released under the Apache License in the same month.
Alibaba released the proprietary Qwen3.7 models in Max and Plus variants in May and June 2026, respectively. The open-source community contributed fine-tuned and abliterated versions, including "Qwable", which incorporated tuning data from Anthropic's Fable 5.
Qwen3.8-Max
Alibaba previewed its 2.4-trillion-parameter model Qwen3.8-Max in July 2026, announcing plans to release its weights. This announcement came a few days after Moonshot AI released its competing Kimi K3 model. The cloud version of Qwen3.8-Max was released on 3 August 2026, using a sparse mixture-of-experts architecture with approximately 95 billion active parameters per forward pass and supporting a context window of up to one million tokens.
On 12 August 2026, Alibaba released the weights as Qwen3.8-2.4T-A95B. This open-weights model omitted certain cloud features, such as image input and a non-thinking mode. Its license required model providers generating more than US$50 million in annual revenue to obtain a commercial license with Alibaba. At that time, it was the second largest and second most powerful open-weights LLM and Chinese LLM, after Kimi K3.
On 14 August 2026, Alibaba released Qwen3.8-27B, which included both the image input and non-thinking mode omitted in the larger release. This model was released under the more permissive Apache 2.0 license.
Ecosystem and Applications
Qwen models have been downloaded more than 40 million times, with over 100 open-weights models released across the family. Fine-tuned versions have been developed by enthusiasts, such as "Liberated Qwen" by San Francisco-based Abacus AI, which responds to user requests without content restrictions.
A local version of Qwen is used by Apple Intelligence in China, integrated with the operating system. In January 2026, an update to the Qwen mobile application connected the chatbot to Alibaba Group's ecosystem, starting with food-service delivery. Plans included allowing users to assign tasks to platforms such as Taobao and Fliggy and helping with errands like phone calls and document processing.
Leadership and Future Direction
Former Qwen AI model division head Lin Junyang resigned in March 2026 after the release of Qwen3.5 and Qwen3.5-Plus, becoming the third executive to leave Alibaba that year. Amid concerns about a potential shift away from research and open-source AI, Alibaba stated it would continue its focus on open source. Later that month, a company announcement revealed the formation of a new AI division, signaling ongoing commitment to the field.