Jump to content

Qwen

From Wikipedia, the free encyclopedia
Qwen
DeveloperAlibaba Cloud
Initial releaseApril 2023; 2 years ago (2023-04)
Stable release
Qwen3-Max /
September 5, 2025; 57 days ago (2025-09-05)
Qwen3-235B-A22B /
July 25, 2025; 3 months ago (2025-07-25)
Qwen3-Next /
September 11, 2025; 51 days ago (2025-09-11)
Repositorygithub.com/QwenLM/Qwen
Written inPython
Operating system
TypeChatbot
LicenseApache-2.0
Qwen Research License
Qwen License
Websitechat.qwen.ai
Qwen
Tongyi Qianwen
Traditional Chinese通義千問
Simplified Chinese通义千问
Literal meaningto comprehend the meaning, [and to answer] a thousand kinds of questions
Transcriptions
Standard Mandarin
Hanyu PinyinTōngyì Qiānwèn

Qwen (also known as Tongyi Qianwen, Chinese: 通义千问; pinyin: Tōngyì Qiānwèn) is a family of large language models developed by Alibaba Cloud. Many Qwen variants are distributed as open‑weight models under the Apache‑2.0 license, while others are served through Alibaba Cloud.[1]

In July 2024, South China Morning Post reported that benchmarking platform SuperCLUE ranked Qwen2‑72B‑Instruct behind OpenAI's GPT‑4o and Anthropic’s Claude 3.5 Sonnet and ahead of other Chinese models.[2]

Models

[edit]

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.[3][4]

The model's architecture was based on the Llama architecture developed by Meta AI.[5][6] In December 2023, it released its 72B and 1.8B models for download, while Qwen 7B weights were released in August.[7][8] 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.[9][10] Qwen2 contains both dense and sparse models.[11]

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.[12][13] QwQ has a 32K token context length and performs better than o1 on some benchmarks.[14]

The Qwen-VL series is a line of visual language models that combines a vision transformer with a LLM.[5][15] Alibaba released Qwen2-VL with variants of 2 billion and 7 billion parameters.[16][17][18]

In January 2025, Qwen2.5-VL was released with variants of 3, 7, 32, and 72 billion parameters.[19] All models except the 72B variant are licensed under the Apache 2.0 license.[20] 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.[21]

Alibaba has released several other model types such as Qwen-Audio and Qwen2-Math.[22] In total, it has released more than 100 open weight models, with its models having been downloaded more than 40 million times.[10] 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]

List of models
Version Release date Ref.
Tongyi Qianwen September 2023 [38]
Qwen-VL August 2023 [39]
Qwen2 June 2024 [10]
Qwen2-Audio August 2024 [40]
Qwen2-VL December 2024 [16]
Qwen2.5 September 2024 [41]
Qwen2.5-Coder November 2024 [42]
QvQ December 2024 [43]
Qwen2.5-VL January 2025 [44]
QwQ-32B March 2025 [45]
Qwen2.5-Omni March 2025 [29]
Qwen3 April 2025 [30]
Qwen3-Coder July 2025 [46]
Qwen3-Max September 2025 [33]
Qwen3-Next September 2025 [47]
Qwen3-Omni September 2025 [37]

See also

[edit]

References

[edit]
  1. ^ Mo, Liam; Hall, Casey (19 September 2024). "Alibaba accelerates AI push by releasing new open-source models, text-to-video". Reuters.
  2. ^ Jiang, Ben (11 July 2024). "Alibaba's open-source AI model tops Chinese rivals, ranks 3rd globally". South China Morning Post. Archived from the original on 4 March 2025. Retrieved 29 November 2024.
  3. ^ Horwitz, Josh; Ye, Josh (11 April 2023). "Alibaba to roll out generative AI across apps". Reuters.
  4. ^ Hall, Casey (13 September 2023). "Alibaba opens AI model Tongyi Qianwen to the public". Reuters.
  5. ^ a b Bai, Jinze; et al. (28 September 2023). "Qwen Technical Report". arXiv:2309.16609 [cs.CL].
  6. ^ "Qwen/techmemo-draft.md". GitHub. August 3, 2023. Archived from the original on March 7, 2025. Retrieved March 5, 2025.
  7. ^ Fan, Feifei (2023-12-01). "Alibaba unveils new Tongyi Qianwen AI language model". global.chinadaily.com.cn.
  8. ^ Ye, Josh (August 3, 2023). "Alibaba rolls out open-sourced AI model to take on Meta's Llama 2". reuters. Archived from the original on 2023-10-10. Retrieved 2024-11-29.
  9. ^ Jiang, Ben (7 June 2024). "Alibaba says new AI model Qwen2 bests Meta's Llama 3 in tasks like maths and coding". South China Morning Post.
  10. ^ a b c Kharpal, Arjun (19 September 2024). "China's Alibaba launches over 100 new open-source AI models, releases text-to-video generation tool". CNBC.
  11. ^ Yang, An; et al. (10 September 2024). "Qwen2 Technical Report". arXiv:2407.10671 [cs.CL].
  12. ^ Dickson, Ben (29 November 2024). "Alibaba releases Qwen with Questions, an open reasoning model that beats o1-preview". VentureBeat. Archived from the original on 18 January 2025. Retrieved 1 December 2024.
  13. ^ 故渊 (2024-11-28). "阿里通义千问 QwQ 登场：开源 AI 推理新王，MATH 测试超 OpenAI o1 模型 - IT之家". ITHome (Chinese website).
  14. ^ Wiggers, Kyle (27 November 2024). "Alibaba releases an 'open' challenger to OpenAI's o1 reasoning model". TechCrunch.
  15. ^ Browne, Ryan (31 December 2024). "Alibaba slashes prices on large language models by up to 85% as China AI rivalry heats up". CNBC.
  16. ^ a b Franzen, Carl (29 August 2024). "Alibaba releases new AI model Qwen2-VL that can analyze videos more than 20 minutes long". VentureBeat.
  17. ^ 沛霖 (2024-08-30). "阿里通义千问推出 Qwen2-VL：开源 2B / 7B 参数 AI 大模型，处理任意分辨率图像无需分割成块". ITHome (Chinese website).
  18. ^ Wang, Peng; Bai, Shuai; Tan, Sinan; Wang, Shijie; Fan, Zhihao; Bai, Jinze; Chen, Keqin; Liu, Xuejing; Wang, Jialin; Ge, Wenbin; Fan, Yang; Dang, Kai; Du, Mengfei; Ren, Xuancheng; Men, Rui; Liu, Dayiheng; Zhou, Chang; Zhou, Jingren; Lin, Junyang (September 18, 2024). "Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution". Cs.CV. arXiv:2409.12191.
  19. ^ "Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL!". Qwen. 2025-01-26. Retrieved 2025-04-28.
  20. ^ "Qwen/Qwen2.5-VL-72B-Instruct · Hugging Face". huggingface.co. 2025-04-28. Retrieved 2025-04-28.
  21. ^ Jiang, Ben (31 December 2024). "Alibaba Cloud cuts AI visual model price by 85% on last day of the year". South China Morning Post.
  22. ^ Franzen, Carl (8 August 2024). "Alibaba claims no. 1 spot in AI math models with Qwen2-Math". VentureBeat.
  23. ^ Mims, Christopher (April 19, 2024). "Here Come the Anti-Woke AIs". WSJ. Archived from the original on April 23, 2024. Retrieved November 29, 2024.
  24. ^ "Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model". Github. 29 January 2025.
  25. ^ Baptista, Eduardo (January 29, 2025). "Alibaba releases AI model it says surpasses DeepSeek". Reuters.
  26. ^ Qwen, Alibaba (February 24, 2025). "QwQ-Max-Preview". X. Archived from the original on 24 February 2025.
  27. ^ "Qwen2.5-VL-32B: Smarter and Lighter". Qwen. 2025-03-24. Retrieved 2025-03-25.
  28. ^ Nikhil (2025-03-24). "Qwen Releases the Qwen2.5-VL-32B-Instruct: A 32B Parameter VLM that Surpasses Qwen2.5-VL-72B and Other Models like GPT-4o Mini". MarkTechPost. Retrieved 2025-03-25.
  29. ^ a b c Dotson, Kyt (27 March 2025). "Alibaba releases new open-source AI model to power intelligent voice applications". SiliconANGLE.
  30. ^ a b Ara Shaikh, Jasmeen (April 28, 2025). "Alibaba unveils advanced Qwen 3 AI as Chinese tech rivalry intensifies". Reuters.
  31. ^ Wiggers, Kyle (28 April 2025). "Alibaba unveils Qwen3, a family of 'hybrid' AI reasoning models". TechCrunch. Archived from the original on 29 April 2025. Retrieved 29 April 2025.
  32. ^ "Qwen3: Think Deeper, Act Faster". Qwen. 2025-04-29. Retrieved 2025-04-28.
  33. ^ a b Bastian, Matthias (2025-09-07). "Alibaba unveils Qwen3-Max-Preview, its largest language model yet". The Decoder. Retrieved 2025-09-13.
  34. ^ "Big news: Introducing Qwen3-Max-Preview..." X (formerly Twitter). Archived from the original on 2025-09-05. Retrieved 2025-09-13.
  35. ^ "Qwen3 Max - API, Providers, Stats". openrouter.ai. Archived from the original on 2025-09-08. Retrieved 2025-09-13.
  36. ^ "Qwen3-Next: Towards Ultimate Training & Inference Efficiency". Qwen Blog. September 10, 2025. Archived from the original on September 11, 2025. Retrieved September 13, 2025.
  37. ^ a b "Qwen/Qwen3-Omni-30B-A3B-Instruct · Hugging Face". huggingface.co. 2025-09-22. Retrieved 2025-09-23.
  38. ^ Jiang, Ben (13 September 2023). "Alibaba opens Tongyi Qianwen model to public as new CEO embraces AI". South China Morning Post.
  39. ^ Kharpal, Arjun (25 August 2023). "Alibaba launches AI model that can understand images and have more complex conversations". CNBC.
  40. ^ 沛霖 (2024-08-13). "阿里通义千问开源 Qwen2-Audio 7B 语音交互大模型：自由互动，无需输入文本". ITHome (Chinese website).
  41. ^ "Alibaba accelerates AI push by releasing new open-source models, text-to-video". Reuters. September 19, 2024.
  42. ^ Nuñez, Michael (12 November 2024). "Qwen2.5-Coder just changed the game for AI programming—and it's free". VentureBeat.
  43. ^ Dotson, Kyt (26 December 2024). "Alibaba announces advanced experimental visual reasoning QVQ-72B AI model". SiliconANGLE.
  44. ^ Wiggers, Kyle (27 January 2025). "Alibaba's Qwen team releases AI models that can control PCs and phones". TechCrunch.
  45. ^ Franzen, Carl (5 March 2025). "Alibaba's new open source model QwQ-32B matches DeepSeek-R1 with way smaller compute requirements". VentureBeat.
  46. ^ "Alibaba rolls out new AI coding model Qwen3-Coder, says it's their most powerful". Computerworld. Retrieved 2025-07-24.
  47. ^ "Qwen/Qwen3-Next-80B-A3B-Instruct · Hugging Face". huggingface.co. 2025-09-11. Retrieved 2025-09-13.
[edit]
Retrieved from "https://en.wikipedia.org/w/index.php?title=Qwen&oldid=1319567898"