Alibaba rolled out Qwen Chat v0.2 and Qwen2.5-1M model

Alibaba has announced the release of Qwen Chat v0.2, an updated version of its AI-powered platform, bringing new features designed to enhance user productivity and creativity. This update marks a significant step forward in Alibaba's development of multimodal AI tools.

The New Feature: Qwen Chat v0.2

Qwen Chat v0.2 integrates three major functionalities:

1. Web Search: Users can now perform real-time web searches directly within the chat interface.
2. Video Creation: The platform enables text-to-video generation, allowing users to create videos based on prompts.
3. Image Generation: Users can generate high-quality images from text descriptions.

These additions complement existing capabilities such as document analysis, artifact creation, and image understanding, making Qwen Chat a versatile tool for both professional and creative tasks.

Alibaba also has announced the release of its latest open-source language models, Qwen2.5-1M, which introduce groundbreaking capabilities in handling extended token contexts and faster processing speeds. These models are part of Alibaba's ongoing efforts to advance natural language processing technology.

The Qwen2.5-1M series includes two models: Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M. Both models are designed to process an unprecedented 1 million tokens in a single context, making them suitable for tasks that require handling large-scale text inputs, such as document summarization, code analysis, and long-form content generation. This represents a significant leap in context length compared to most existing language models, which typically support token limits in the tens of thousands.

To complement the new models, Alibaba has also open-sourced an advanced inference framework based on vLLM (a high-performance serving system for large language models). This framework integrates sparse attention methods, enabling the models to process 1-million-token inputs with speeds 3 to 7 times faster than traditional approaches. Sparse attention optimizes how the model focuses on relevant parts of the input text, reducing computational overhead while maintaining accuracy.

The release includes detailed documentation in a technical report and blog post, providing insights into the architecture and performance of the Qwen2.5-1M series. Users can explore these models through various platforms: Alibaba’s Qwen Chat for live interaction, Hugging Face for experimentation, and Modelscope for additional deployment options.

About Qwen

Alibaba Cloud developed Qwen Chat as part of its broader AI initiative to create accessible, high-performing tools. The Qwen model family includes various specialized models, such as Qwen2.5-Coder for programming and Qwen2-VL-Max for vision-language tasks. These models are known for their robust multilingual support and extended context length capabilities, with some models handling up to 128K tokens.

When and Where?

Qwen Chat v0.2 was officially rolled out on January 26, 2025, as announced on Alibaba's official X (formerly Twitter) account. The platform is accessible via its dedicated web interface.

How It Works

Users can access these features through toggles in the chat interface:

• Web search results are integrated into conversations.
• Text-to-video and image generation tools allow direct input of prompts to create media outputs.
• Existing functionalities like document uploads and artifact creation remain available within the same interface.

This update represents Alibaba's continued commitment to advancing AI technology while preparing for further developments around the Chinese Spring Festival.