TLDR
- Alibaba Cloud unveiled over 100 new open source large language models under its Qwen 2.5 family of models.
- It also took the wraps off a new text-to-video model that takes Chinese and English text prompts.
- Alibaba Cloud updated its full-stack AI infrastructure as well, to offer powered-up capabilities.
The cloud division of Chinese e-commerce giant Alibaba has unveiled over 100 new large language models (LLMs) from its Qwen 2.5 family of models.
The Qwen 2.5 series, ranging in size from 500 million to 72 billion parameters, offers models that are tailored for enhanced knowledge, math, coding, and support for over 29 languages. These open-source models are now accessible to developers and researchers across sectors like gaming, automotive, and scientific research, enabling applications both at the edge and in the cloud.
“Today marks a significant milestone as we launch our most expansive open-source initiative to date,” said Alibaba Cloud CTO Jingren Zhou in a blog post.
Since first launching in April 2023, the Qwen models have garnered over 40 million downloads across platforms such as Hugging Face and Alibaba’s ModelScope. The Qwen 2.5 release includes a wide array of base models, instructional models, and specialized ones for tasks like language, vision, audio, and code applications.
Multimodal advancements and Qwen-Max
Alibaba Cloud also announced an enhanced version of its flagship model, Qwen-Max, which demonstrates superior performance in key AI tasks, including math, coding, and reasoning.
The company also unveiled a new text-to-video model within the Tongyi Wanxiang large model family. Capable of generating high-quality videos in various styles, this model uses Chinese and English text to convert static images into videos using advanced diffusion transformer (DiT) architecture.
In the multimodal AI frontier, Alibaba updated its vision language model with the new Qwen2-VL, which brings advanced reasoning capabilities that can comprehend and interact with videos up to 20 minutes long, supporting applications in mobile devices, automobiles, and robotics.
It also unveiled AI Developer, an AI-powered assistant that automates coding tasks such as bug fixing, requirement analysis, and code programming, allowing developers to focus on high-priority duties.
AI infrastructure overhaul
Alibaba Cloud also updated its AI infrastructure, designed to bolster the development and deployment of AI models. The revamped full-stack infrastructure includes the following:
- CUBE DC 5.0 Data Center Architecture: This next-gen architecture is optimized for AI development, with energy-efficient innovations like wind-liquid hybrid cooling systems and smart management tools that cut deployment times by up to 50%.
- Open Lake Solution: A new data management system aimed at maximizing the utility of generative AI by integrating big data engines and achieving cost savings through enhanced compute-storage separation.
- PAI AI Scheduler: A cloud-native scheduling engine that achieves over 90% compute utilization by intelligently managing AI model training and inference tasks.
