Following the global attention generated by the open-source release of DeepSeek R1, more large model companies are accelerating the launch of deep thinking models, with plans to open-source them or having already done so.
Alibaba's Tongyi team recently launched a new deep thinking model, QwQ-Max-Preview, which supports online search. This model is currently available on the new Tongyi Qianwen website (https://chat.qwen.ai/) and will be open-sourced soon.
QwQ-Max-Preview Overview
QwQ-Max-Preview is an inference model built upon Qwen2.5-Max. As Alibaba's important ultra-large-scale MoE model, Qwen2.5-Max has been trained on over 20 trillion tokens and demonstrates outstanding performance in various benchmark tests, surpassing leading AI models like DeepSeek V3 and GPT-4o, laying a solid foundation for QwQ-Max-Preview.
Currently in preview, the Alibaba Qwen team has stated that it will continue to optimize the model. A formal release of QwQ-Max is planned, along with Android and iOS apps. The weights of QwQ-Max and Qwen2.5-Max will be open-sourced under the Apache2.0 license. Smaller models, such as QwQ-32B, deployable on local devices, are also planned for release.
It supports two core functionalities: deep thinking and online search. The deep thinking capability allows for in-depth analysis of complex problems, while the online search function enables the model to overcome its knowledge limitations by accessing real-time internet information for reasoning.
QwQ-Max-Preview Highlights
- Powerful Mathematical Understanding: QwQ-Max-Preview excels at tackling various mathematical challenges, from basic calculations to complex mathematical competition problems. Its deep thinking function allows it to dissect problem-solving approaches and provide accurate solutions. For example, when solving for function extrema in advanced mathematics, it can thoroughly analyze the function's properties and systematically derive the extreme points.
- Exceptional Programming Capabilities: The model can handle tasks ranging from creating simple scripts to developing complex applications. For instance, given the prompt "Create a p5.js script to generate a rotating sphere containing 100 bouncing yellow balls," it can output detailed code with functional explanations.
- Deep Reasoning and Logical Analysis: It demonstrates rigorous reasoning in handling logical problems. For example, in the problem "A father and daughter's ages add up to 100 years, the father is 50, how old is the daughter?", it not only calculates the mathematical answer but also analyzes the illogical aspects from a realistic perspective, suggesting possibilities such as adoption or science fiction elements to make the problem's logic more complete.
- Online Search and Information Integration: It can quickly access internet information and reason based on real-time data. When evaluating the movie "Ne Zha 2," it can quickly search for relevant information, clearly citing sources in its reasoning and answer, integrating information from multiple sources to provide a comprehensive evaluation.
- Multimodal Capabilities Planned: While not fully supported in the current preview version, related functionalities are planned. For example, given the prompt "Generate an image of a girl wearing a white dress, then display a QR code linking to taobao.com," it can generate an image in a chosen style (e.g., anime) along with a functional QR code.
Application Scenarios
- Education and Learning: Students can use QwQ-Max-Preview's deep thinking and problem-solving abilities to get detailed solutions and approaches to problems in mathematics, programming, and other subjects. For example, when learning algorithmic programming, they can ask the model to help analyze code logic errors or provide optimization suggestions.
- Creative Work: Writers, designers, and other creative professionals can use the online search function to get inspiration. For example, writers of science fiction can use the model to search for cutting-edge scientific information to add scientific accuracy and imagination to their works; designers can have the model generate design sketches or creative concepts.
- Game Development: Game developers can leverage the model's capabilities to design new gameplay, such as combining Tetris and 2048. The model can provide creative solutions and generate related code to some extent. Although the preview version's output may have inaccuracies, it can still provide developmental ideas.
- Daily Problem Solving: Ordinary people can use the model to solve various problems in their daily lives. For example, when planning a trip, they can use the model to search for travel guides and attraction information; when repairing appliances, they can search for troubleshooting and repair methods.
- Industry Research: Researchers and analysts can use the model to search and integrate industry news and cutting-edge technologies to assist in writing research reports and analyzing industry trends.
QwQ-Max-Preview User Guide
- Access the Platform: Users can access the Qianwen chat platform via chat.qwen.ai.
- Enable Features: In the chat box, find the "Deep Thinking (QwQ)" button and the "Online Search" option in the lower left corner. Click the "Deep Thinking (QwQ)" button to enable the deep thinking function; to use online search to assist reasoning, also enable the "Online Search" option.
- Ask Questions: Accurately and clearly describe the question in the input box, such as "Analyze the current development trend of the new energy vehicle industry" or "Write a Python code to implement data sorting."
- View Results: The model will think and reason, and if online search is enabled, it will integrate online information and provide an answer. The answer may include text analysis, code, images (if multimodal functions are involved), etc. If you are not satisfied with the results, you can refine the question and ask again.
Conclusion
Alibaba's QwQ-Max-Preview deep thinking model, with its outstanding capabilities in mathematics, programming, reasoning, and other areas, as well as its unique online search and multimodal layout, opens a new door to AI experience for users. Although currently in the preview stage, there may be some minor flaws in actual use, but with the subsequent release of the official version and continuous optimization, it is believed that it will play a greater role in many fields such as education, creativity, and development.
Welcome to share your usage experience and suggestions in the comments section, and witness the continuous growth and improvement of this model. We also look forward to Alibaba's continued innovation in the AI field, bringing us more surprises and breakthroughs.
