Alibaba Cloud, the digital technology and intelligence backbone of Alibaba Group, recently announced it has released over 100 of its newly launched large language models, Qwen 2.5, to the global open-source community. This significant contribution was revealed at the Apsara Conference, its annual flagship event.
In addition, Alibaba Cloud has unveiled a revamped full-stack infrastructure designed to meet the growing demands for robust AI computing. This new infrastructure includes innovative cloud products and services that enhance computing, networking, and data center architecture, all aimed at supporting the thriving development and wide-range applications of AI models.
“Alibaba Cloud is investing, with unprecedented intensity, in the research and development of AI technology and the building of its global infrastructure. We aim to establish an AI infrastructure of the future to serve our global customers and unlock their business potential,” said Eddie Wu, Chairman and Chief Executive Officer of Alibaba Cloud Intelligence.
100 Open-sourced Models Unveiled
The newly released open-source Qwen 2.5 models, ranging from 0.5 to 72 billion parameters in size, feature enhanced knowledge and stronger capabilities in math and coding and are able to support over 29 languages, catering to a wide array of AI applications both at the edge or in the cloud across various sectors from automobile, gaming to science research.
The Qwen model series, Alibaba Cloud’s portfolio of proprietary large language models, has achieved remarkable traction since its debut in April 2023. To date, the Qwen models have surpassed 40 million downloads across platforms such as Hugging Face and ModelScope, an open-source community initiative by Alibaba. Furthermore, these models have inspired the creation of over 50,000 models on Hugging Face.
The Qwen 2.5 release will see over 100 models being made open-source. This extensive range includes base models, instruct models, and quantized models of various precision levels and methods, spanning different modalities such as language, audio, and vision, along with specialized code and mathematical models.
“Today marks a significant milestone as we launch our most expansive open-source initiative to date,” said Jingren Zhou, Chief Technology Officer of Alibaba Cloud Intelligence. “This initiative is set to empower developers and corporations of all sizes, enhancing their ability to leverage AI technologies and further stimulating the growth of the open-source community. We remain committed to investing in advanced AI infrastructure to foster the widespread adoption of generative AI technologies across different industries.”
Alibaba Cloud also announced an upgrade to its proprietary flagship model Qwen-Max. The enhanced Qwen-Max model demonstrates performance on par with other state-of-the-art models in areas such as language comprehension and reasoning, math, and coding.
Expanding the Frontier in Multimodal
In addition to its extensive suite of large language models, Alibaba Cloud also unveiled a new text-to-video model as part of its image generator, Tongyi Wanxiang large model family. The new model is capable of generating high-quality videos in a wide variety of visual styles from realistic scenes to 3D animation. The model can generate videos based on Chinese and English text instruction and transform static images into dynamic videos. The model features advanced diffusion transformer (DiT) architecture to enhance video reconstruction quality.
The cloud leader is also deploying a significant update to its vision language model with the introduction of Qwen2-VL, capable of comprehending videos lasting over 20 minutes and support video-based question-answering. Equipped with sophisticated reasoning and decision-making capabilities, Qwen2-VL is designed for integration into mobile phones, automobiles and robots, facilitating the automation of specific operations.
For computer programming, Alibaba Cloud has also launched an AI Developer, a Qwen-powered AI assistant designed to support programmers automate tasks such as requirement analysis, code programming and identifying and fixing software bugs. This enables developers to concentrate more on essential duties and further their skills.
A Full-Stack AI Infrastructure Upgrade
The cloud pioneer has also announced a slew of innovative updates to its full-stack AI infrastructure covering green datacenter architecture, data management, model training and inferencing:
- Next-Gen Data Center Architecture for Surging AI Development: To meet the increasing and diverse demand for high-performance computing power driven by the global AI boom, Alibaba Cloud has revealed its next-generation data center architecture, CUBE DC 5.0. The new CUBE architecture increases energy and operational efficiency with a set of advanced and proprietary technologies such as wind-liquid hybrid cooling system, all-direct current power distribution architecture and smart management system and reduces deployment times by up to 50% compared to traditional data center builds through prefabricated modular designs.
- Open Lake Solution to Maximize Data Utility: As organizations face challenges in managing vast amounts of data amidst the growing demand for generative AI, Alibaba Cloud introduces Alibaba Cloud Open Lake which can seamlessly integrate big data engines into a unified solution, maximizing data utility especially for generative AI applications. By integrating workflows, performance optimization, and robust governance in a single platform, it achieves efficient resource usage through compute-storage separation, clear data governance, and significant cost and time savings.
- AI Scheduler with Integrated Model Training and Inference: Alibaba Cloud has launched PAI AI Scheduler with integrated model training and inference, a proprietary cloud-native scheduling engine designed to enhance computing resource management. Through utilizing intelligent integration of diverse computing resources, flexible resource scheduling, real-time tasks adjustments, and automatic fault recovery, it can achieve over 90% of effective compute utilization rate.
- DMS for Unified Management of Metadata: To help organizations efficiently manage their data and unleash values, Alibaba Cloud introduced DMS: OneMeta+OneOps, a platform that enables a unified management of over 40 types of data sources in database, data warehouse, and data lake across multiple cloud environments. The platform will boost data utilization rate by 10 times, significantly enhancing the efficiency of transforming data into valuable intelligence.
- More Powerful Elastic Compute Service: Alibaba Cloud also introduced the 9th Generation Enterprise Elastic Compute Service (ECS) instance during the conference. The latest generation of ECS instances has notable performance enhancements, including a 30% increase in search recommendation speed and a 17% improvement in the effectiveness of reading and writing Queries Per Second (QPS) when applying to database products compared to the previous generation.
These updates are designed to provide more comprehensive support for customers and partners to maximize the benefits of the latest technology has to offer for building even more efficient, sustainable and inclusive AI applications.
