Qwen
A family of large and small language models developed by Alibaba Cloud.
Qwen is a series of open-weights and proprietary language models created by Alibaba Cloud. Starting in April 2023, the family has expanded to include multimodal and reasoning variants, with the latest version reaching 2.4 trillion parameters by August 2026.
- Released by Alibaba Cloud starting in April 2023.
- Latest version Qwen3.8 reached 2.4 trillion parameters.
- Supports text, vision, audio, and video modalities.
- Uses sparse mixture-of-experts architecture.
- Available under Apache 2.0 or proprietary licenses.
What it is
Qwen, also known as Tongyi Qianwen, is a family of large and small language models developed by Alibaba Cloud. The project launched its first beta in April 2023 and became publicly available later that year. It includes dense and sparse architectures, ranging from small on-device models to massive systems with trillions of parameters. The latest iteration, Qwen3.8, features a 2.4 trillion parameter model released in August 2026. These models support text, vision, audio, and video inputs and outputs.
How it came about
The initial version relied on Meta AI's Llama architecture and released weights for its 72B model in December 2023. Subsequent releases like Qwen2 in June 2024 introduced both dense and sparse models. By early 2025, the family expanded to include visual and omnimodal variants. In April 2025, Qwen3 arrived with Apache 2.0 licensing. The timeline continued through 2026 with versions 3.5, 3.6, 3.7, and finally 3.8, which marked a significant increase in scale and capability for the open-weight community.
Inside the model
The models utilize decoder-only transformer architectures, including sparse mixture-of-experts designs. The flagship Qwen3.8-Max uses approximately 95 billion active parameters per forward pass despite having 2.4 trillion total parameters. Training involved 36 trillion tokens across 119 languages. The system supports a context window of up to one million tokens. Architectures vary between dense models for smaller sizes and MoE configurations for larger ones, allowing efficient inference while maintaining high capacity.
What it can do
Qwen models handle text generation, image understanding, audio processing, and video analysis. The Qwen3 family introduced a switchable thinking mode for complex reasoning tasks. Benchmarks indicate strong performance, with variants like QwQ-32B-Preview outperforming competitors on specific tests. The largest open-weight model, Qwen3.8-2.4T-A95B, was noted as the second most powerful open LLM by mid-August 2026. Multimodal capabilities allow real-time voice chatting and comprehensive document analysis.
How to use it
Access varies by license. Many models are available under the Apache 2.0 license, permitting broad use and modification. Larger proprietary models are accessible via API through Alibaba Cloud. Some versions require revenue-sharing agreements if annual earnings exceed US$50 million. Weights are distributed on platforms like Hugging Face and ModelScope. Smaller distilled versions often carry more permissive terms than their massive counterparts.
Safety and testing
Models must comply with Chinese regulations regarding content moderation. This includes restrictions on topics deemed sensitive by local authorities. Safety evaluations are part of the development process, though specific red-teaming details are not fully disclosed in public sources. The integration into consumer apps involves filtering mechanisms to ensure compliance with national standards for generative AI services.
Reception and debate
The open-source strategy has garnered significant adoption, with over 200,000 variations listed on Hugging Face. The framework is viewed as critical for China's AI development, helping overcome compute constraints. However, it faces scrutiny regarding censorship protocols required by the Cyberspace Administration of China. International users appreciate the open weights, while domestic deployment ensures alignment with local legal frameworks.
Why it matters
Qwen has become a foundational tool for many open-source projects and fine-tuned derivatives. Its availability has influenced regional AI strategies, such as Apple Intelligence integrating a local Chinese variant. The release of large open-weight models has shifted competition in the global LLM landscape. Developers frequently use Qwen as a base for specialized applications in e-commerce, research, and robotics sectors.
Questions people ask about Qwen
Who developed Qwen?
Qwen was developed by Alibaba Cloud, a subsidiary of Alibaba Group. The project began as a research initiative within the company's Tongyi Laboratory before becoming a major product line.
Is Qwen open source?
Many Qwen models are open-weights and licensed under Apache 2.0, allowing free use and modification. However, some advanced variants remain proprietary and are only accessible via API or require commercial licensing agreements.
What is the largest Qwen model?
As of August 2026, the largest model is Qwen3.8-Max with 2.4 trillion parameters. The open-weight version, Qwen3.8-2.4T-A95B, activates approximately 95 billion parameters during inference.
Does Qwen support multiple languages?
Yes, the models were trained on data covering 119 languages and dialects. They are designed to understand and generate text in multilingual contexts effectively.
Good reading and sources
- Official websiteOfficial site
- China's Alibaba launches over 100 new open-source AI models, releases text-to-video generation toolCNBC, Article
The facts on this page come from the Wikipedia article "Qwen", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-15.