Qwen
Qwen, developed by Alibaba Cloud, became the top Chinese language model in July 2024, ranking third in the world behind models from Anthropic and OpenAI.
Key features
Qwen supports switching between various models. Each is good for some particular tasks.
The main features include:
- Artifacts - reusable content templates or components that make repetitive tasks easier. These artifacts can be anything from standardized document formats to code snippets. This feature is especially useful for automating routine processes, but only if the instructions are clear and the output is complex enough.
- Web Search - finds up-to-date information from reliable online sources. This feature lets users get real-time insights without leaving the platform, making research easier.But the accuracy and relevance of the results depend on the quality of the online resources and how well the search is done. It's especially useful for people who need quick access to current trends, news, or factual details, but it might not replace in-depth research done through specialized databases or with expert help.
- Image generation - uses advanced algorithms to do this. Users can request different types of images, from simple illustrations to complex designs. The tool is very good at creating images that are visually appealing. And it really does a good job!
- Video generation - can create up to 5 seconds long clips (for now).
Generated image example.
Prompt: сreate an image of a beautiful sunset on the shore of a beautiful tropical beach, the scene is built so that the frame includes the ocean, beautiful clouds illuminated by the sun rays in different colors of bright red, pink, orange, yellow and blue, as well as the shore, palm trees and a small mountain. Several birds fly in the sky, waves crash on the shore, and a light breeze blows. HD quality, ultra-realistic Canon photography.
Development
- April 2023: Alibaba launched Qwen (Tongyi Qianwen) based on Meta AI's Llama, with modifications.
- September 2023: Qwen received government approval and went public.
- December 2023 & August 2024: Alibaba released several models, including 72B, 1.8B, and 7B versions, as open source.
- June 2024: Qwen 2 was launched, combining various experts.
- September 2024: some models were open-sourced, while advanced models remained proprietary.
- November 2024: QwQ-32B-Preview, focusing on reasoning, was released under the Apache 2.0 License.
- January 2025: Alibaba launched Qwen 2.5-Max, outperforming models such as GPT-4o, DeepSeek-V3, and Llama-3.1-405B in key benchmarks.
The Qwen-VL series combines vision transformers with language models. Alibaba released Qwen-VL2 with 2 billion and 7 billion parameter models. Qwen-VL-max, Alibaba's flagship vision model, costs $0.00041 per thousand input tokens.
Alibaba has also launched Qwen-Audio and Qwen2-Math. More than 100 models have been released as open source, with more than 40 million downloads. Enthusiasts have created versions such as "Liberated Qwen," which removes content restrictions.
Pricing (has free plan)
As for January 2025, Qwen is still free to use.
Summary
Designed to generate human-like responses that balance technical precision with conversational tone. While it works well across a wide range of topics, its adaptability to highly specialized or niche topics may be limited compared to domain-specific tools or experts. It is best suited for content creators, marketers, and writers who need engaging, SEO-friendly material for general audiences, but may not meet the needs of users who require deep expertise in narrowly defined fields.
FAQ
Qwen is a large language model developed by Alibaba Cloud. It can assist with tasks like answering questions, writing, and coding. It is multilingual and has advanced reasoning capabilities.
Yes, it can.
It makes sure your data is safe and private by using encryption, following global standards, and limiting who can access it. Your data is not stored or shared without your permission.
Yes (January 2025). In future they will probably introduce some paid plans with advanced features.
Yes, Qwen supports fine-tuning methods like SFT (Supervised Fine-Tuning) and RLHF (Reinforcement Learning from Human Feedback) for specific use cases.
Qwen competes with models like GPT-4o and Llama-3.1-405B in areas like reasoning, multilingual support, and efficiency. Benchmarks show strong performance across various tasks.
Some versions of Qwen are open source, including smaller models like Qwen-7B. Access them via platforms like ModelScope or GitHub.
Yes, Qwen excels at reasoning and problem-solving, especially with specialized variants like QwQ (Qwen with Questions), which enhances logical and multi-step reasoning.
Qwen supports multiple languages, including Chinese, English, French, Spanish, Portuguese, Russian, Arabic, Japanese, Korean, Vietnamese, Thai, Indonesian, and more.
Qwen tries to be accurate by using a lot of training data and complex algorithms. But users should check important information because AI-generated content can have limits and even hallucinations.
Yes, Qwen and its variants can be used for commercial projects, subject to licensing agreements. Check Alibaba Cloud’s terms for specific use cases.
Can hallucinate and forget instructions.
I like "Artifact" feature, which allows you to preview the code of landing page, for example, in real time is great. But again this feature has bugs, when it cannot reproduce some code features, eg it could not show me a simple alert message. But the code was correct, its just some rendering problem I think, which will be fixed.
And I like it's image generation. It's not Midjourney, it's not great, but good and fast enough for some basic stuff.