Top Qs
Timeline
Chat
Perspective
Qwen
Family of large language models by Alibaba From Wikipedia, the free encyclopedia
Remove ads
Qwen (also called Tongyi Qianwen, Chinese: 通义千问) is a family of large language models developed by Chinese company 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]
Remove ads
Models
Summarize
Perspective
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 for download, while Qwen 7B weights were released in August.[6][7] Their models are sometimes described as open source, but the training code has not been released nor has the training data been documented, and they do not meet the terms of either the Open Source AI Definition or the Model Openness Framework from the Linux Foundation.
In June 2024 Alibaba launched Qwen2 and in September it released some of its models with open weights, while keeping its most advanced models proprietary.[8][9] Qwen2 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$0.41 per million 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 weight 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]
On January 29, 2025, Alibaba launched Qwen2.5-Max. According to a blog post from Alibaba, Qwen2.5-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 opened up, 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 Qwen3 model family was released,[30] with all models licensed under the Apache 2.0 license. The Qwen3 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 Qwen3 models support reasoning, which can be enabled or disabled through the tokenizer. The Qwen3 models are available through chat.qwen.ai and can be downloaded via Hugging Face and ModelScope.[32]
On September 5, 2025, Alibaba launched Qwen3-Max.[33] According to Alibaba's official X account, it outperforms other foundation non-reasoning models such as Qwen3-235B-A22B-Instruct-2507, Kimi K2, Claude 4 Opus Non-thinking, and DeepSeek V3.1.[34] There is no dedicated thinking mode for Qwen3-Max as of yet.[35]
On September 10, 2025, Qwen3-Next was released under the Apache 2.0 license and made available through chat.qwen.ai, as well as platforms like Hugging Face and Model Scope. Qwen3-Next includes two post-trained Instruct and Thinking models. Qwen3-Next was created with a new model-architecture called Qwen3-Next, in the belief that Context Length Scaling and Total Parameter Scaling are two major trends in the future of large models. Qwen3-Next introduces several key improvements over the Qwen3 architecture: a hybrid attention mechanism, a highly sparse mixture-of-experts (MoE) structure, training-stability-friendly optimizations, and a multi-token prediction mechanism for faster inference. Based on the Qwen3-Next architecture, a model with 80B total parameters and 3B active parameters was created. The Qwen3-Next model performs comparable to, or in some cases better than, Qwen3-32b while using less than 10% of its training cost (in GPU hours). In inference, especially with contexts greater than 32K tokens, it reaches greater than 10x higher throughput. Qwen3.5 will use a refined version of the Qwen3-Next architecture.[36]
On September 22, 2025, Qwen3-Omni was release under the Apache 2.0 license and made available through chat.qwen.ai, as well as platforms like Hugging Face and Model Scope. Qwen3-Omni is a mixed/multimodal model that can process text, images, audio, and video, and deliver real-time streaming responses in both text and natural speech.[37]
Remove ads
See also
References
External links
Wikiwand - on
Seamless Wikipedia browsing. On steroids.
Remove ads