Qwen

Qwen

Alibaba launched a beta of Qwen in April 2023 under the name Tongyi Qianwen, then opened it for public use in September 2023 after regulatory clearance.^[4][5]

The model's architecture was based on the Llama architecture developed by Meta AI.^[6][7] In December 2023, it released its 72B and 1.8B models for download, while Qwen 7B weights were released in August.^[8][9]

Qwen2

Qwen2 was released in June 2024, and in September it released some of its models with open weights, while keeping its most advanced models proprietary.^[10][11] Qwen2 contains both dense and sparse models.^[12]

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.^[13][14] QwQ has a 32K token context length and performs better than o1 on some benchmarks.^[15] It was also in November 2024 that the Accio application was launched.^[16] Accio is an AI native application that is built upon Qwen and is used to generate market insights and answer sourcing questions for Alibaba's business to business e-commerce site. The tool is able to automate labor intensive tasks like data collection and trend tracking. ^[17]

The Qwen-VL series is a line of visual language models that combines a vision transformer with an LLM.^[6][18] Alibaba released Qwen2-VL with variants of 2 billion and 7 billion parameters.^[19][20][21]

In January 2025, Qwen2.5-VL was released with variants of 3, 7, 32, and 72 billion parameters.^[22] All models except the 72B variant are licensed under the Apache 2.0 license.^[23] Qwen-VL-Max is Alibaba's flagship vision model as of 2024, and is sold by Alibaba Cloud at a cost of US$0.41 per million input tokens.^[24]

Alibaba has released several other model types such as Qwen-Audio and Qwen2-Math.^[25] In total, it has released more than 100 open weight models, with its models having been downloaded more than 40 million times.^[11] 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.^[26]

On January 29, 2025, Alibaba launched Qwen2.5-Max.^[27][28]

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.^[29][30]

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. 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.^[31]

Qwen3

On April 28, 2025, the Qwen3 model family was released,^[32] with all models licensed under the Apache 2.0 license. The Qwen3 model family includes both dense and MoE models. The sizes of the dense models include 0.6B, 1.7B, 4B, 8B, 14B, and 32B, and the MoE models include 30B-A3B (30B with 3B activated parameters) and 235B-A22B (235B with 22B activated parameters)^[33]. They were trained on 36 trillion tokens in 119 languages and dialects.^[34]

The Qwen3 collection include^[33]:

• Qwen3, which supports switching between thinking and non-thinking mode
• Qwen3 Base, the pretrained base model
• Qwen3 Instruct, which supports only non-thinking mode
• Qwen3 Thinking, which supports only thinking mode

In additional to open-weights models, Qwen3 also includes Qwen3-Max, a proprietary large model with over 1T parameters, trained with about 36T tokens, that is available through API,^[35] and Qwen3-Max-Thinking, its reasoning variant that can generate text, pictures, or video.^[36]

List of models