Alibaba rolled out a free way to test drive its Qwen AI models through a new web interface called Qwen Chat.
The platform lets users experiment with several AI models side by side, including Qwen2.5-Plus for general conversation, Qwen2-VL-Max for understanding images, and specialized models like QwQ and QVQ for reasoning tasks.
For developers, the interface includes access to Qwen2.5-Coder-32B-Instruct, a model designed for programming tasks. Users can upload their own documents and receive responses based on that content. The interface also includes a preview mode with HTML support and the ability to process uploaded images for visual analysis.
Video: Qwen
Alibaba plans to expand the platform's capabilities with web search, image generation capabilities, and voice interaction mode. Like other AI models developed in China, the Qwen models include content filtering that restricts content critical of the government.
Alibaba's growing family of models
Alibaba's cloud division has been steadily building its AI capabilities since introducing the first Qwen models in August 2023. The newer Qwen2 series brought notable improvements across several areas, including programming, mathematics, logical reasoning, and support for multiple languages.
The current Qwen 2.5 series comes in three flavors: a general-purpose language model, Qwen2.5-Coder for programming tasks, and Qwen2.5-Math for mathematical problems. Most of these models are available as open-source software under the Apache 2.0 license.
The team recently expanded into AI agents with their new Qwen-Agent framework, which builds on their existing language models. The system can follow instructions, work with various tools, plan complex tasks, and maintain context during conversations. It also includes some advanced features, like RAG integration and a specialized math reasoning system powered by Qwen2.5-Math.
Their latest experiment, QwQ-32B-Preview, takes a page from OpenAI's playbook by focusing on test-time compute scaling for better logical reasoning and problem-solving abilities. The model follows similar principles to OpenAI's o1 approach to improve its performance.
