Qwen (also called Tongyi Qianwen, Chinese: 通义千问) is a family of large language models developed by Alibaba Cloud. In July 2024, it was ranked as the top Chinese language model in some benchmarks and third globally behind the top models of Anthropic and OpenAI.[1]
Models
Alibaba first launched a beta of Qwen in April 2023 under the name Tongyi Qianwen.[2] The model's architecture was based on the Llama architecture developed by Meta AI.[3][4] It was publicly released in September 2023 after receiving approval from the Chinese government.[5] In December 2023 it released its 72B and 1.8B models as open source, while Qwen 7B was open sourced in August.[6][7]
In June 2024 Alibaba launched Qwen 2 and in September it released some of its models as open source, while keeping its most advanced models proprietary.[8][9] Qwen 2 contains both dense and sparse models.[10]
In November 2024, QwQ-32B-Preview, a model focusing on reasoning similar to OpenAI's o1, was released under the Apache 2.0 License, although only the weights were released, not the dataset or training method.[11][12] QwQ has a 32K token context length and performs better than o1 on some benchmarks.[13]
The Qwen-VL series is a line of visual language models that combines a vision transformer with a LLM.[3][14] Alibaba released Qwen2-VL with variants of 2 billion and 7 billion parameters.[15][16][17]
In January 2025, Qwen2.5-VL was released with variants of 3, 7, 32, and 72 billion parameters.[18] All models except the 72B variant are licensed under the Apache 2.0 license.[19] Qwen-VL-Max is Alibaba's flagship vision model as of 2024 and is sold by Alibaba Cloud at a cost of US$0.00041 per thousand input tokens.[20]
Alibaba has released several other model types such as Qwen-Audio and Qwen2-Math.[21] In total, it has released more than 100 open source models, with its models having been downloaded more than 40 million times.[9][22]Fine-tuned versions of Qwen have been developed by enthusiasts, such as "Liberated Qwen", developed by San Francisco-based Abacus AI, which is a version that responds to any user request without content restrictions.[23]
In January 2025, Alibaba launched Qwen 2.5-Max. According to a blog post from Alibaba, Qwen 2.5-Max outperforms other foundation models such as GPT-4o, DeepSeek-V3, and Llama-3.1-405B in key benchmarks.[24][25] In February 2025, Alibaba announced on their official X account that the 2.5-Max model would be open-sourced, however it has not been released.[26]
On March 24 2025, Alibaba launched Qwen2.5-VL-32B-Instruct as a successor to the Qwen2.5-VL model. It was released under the Apache 2.0 license.[27][28]
On March 26, 2025, Qwen2.5-Omni-7B was released under the Apache 2.0 license and made available through chat.qwen.ai, as well as platforms like Hugging Face, GitHub, and ModelScope.[29] The Qwen2.5-Omni model accepts text, images, videos, and audio as input and can generate both text and audio as output, allowing it to be used for real-time voice chatting, similar to OpenAI's GPT-4o.[29]
On April 28, 2025, the Qwen 3 model family was released,[30] with all models licensed under the Apache 2.0 license. The Qwen 3 model family includes both dense (0.6B, 1.7B, 4B, 8B, 14B, and 32B parameters) and sparse models (30B with 3B activated parameters, 235B with 22B activated parameters). They were trained on 36 trillion tokens in 119 languages and dialects.[31] All models except the 0.6B, 1.7B, and 4B variants have a 128K token context window. Like OpenAI's o1 and QwQ 32B, the Qwen 3 models support reasoning, which can be enabled or disabled through the tokenizer. The Qwen 3 models are available through chat.qwen.ai and are open-sourced on Hugging Face and ModelScope.[32]
