This week, Alibaba announced cutting edge AI advances across enterprise security, data integration frameworks and spatial intelligence. Highlights include the release of Model Studio: Exclusivea hybrid and private cloud AI development platform for businesses. Web Shapera pioneering data integration method for training research agents with a precise focus. And upgrade Qwen-driven AMAP travel agencythe world's first AI-Native navigation tool that can handle complex travel queries. Together, these innovations underscore Alibaba's willingness to rebuild the industry through safe, scalable, and hyper-intelligent AI solutions.
Alibaba Cloud announces “Model Studio: Exclusive” at Indonesia AI Conference 2025
Alibaba Cloud recently launched an exclusive version of its enterprise-grade AI development platform at Model Studio: Exclusive, Indonesia AI Conference 2025. This platform is an ideal option for organizations that require both innovation and high security.
Private clouds are growing demand, especially in finance, healthcare and public services, as organizations prioritize security, compliance, AI and customization.
So Alibaba Cloud deployed Model Studio: Exclusive, it Designed for hybrid and private cloud environments, it helps businesses make the most of the possibilities of large-scale language models (LLMs) and intelligent AI agents without risking the safety of their infrastructure.
From the basic models of fine tuning to building industry-specific AI agents, the platform helps businesses deploy tailored AI solutions in a secure environment. Key features include analyzing smart data and safely extracting insights in a high-compliance environment. Automated data integration and intelligent quality assurance to streamline your training workflow. Cross-modality data processing to optimize post-training. It also supports end-to-end task automation, where AI agents can break down complex tasks, plan the next step, and complete tasks using a variety of tools.
Alibaba announces innovative training frameworks to improve deep research agents
The lack of high quality data poses a broad challenge in training deep research agents. To improve training performance, Alibaba introduced Webshaper, a pioneering data synthesis framework dedicated to overcoming limitations.
Deep Research Agents are AI agents designed to autonomously collect, retrieve and process information from a variety of sources. Their performance depends heavily on the complexity and quality of the training data. A typical data synthesis method follows an information-driven approach, extracting web data and creating relevant questions and answers. This method can lead to discrepancies between extracted data, generated questions and answers.
WebShaper offers a solution with a formalization-driven approach that transforms queries and information-seeking tasks into accurate mathematical representations. Start data integration by building a “seed” task, increasing the complexity of these tasks multiple times using search and validation tools based on mathematical formalization.
This new approach helps WebShaper help AI agents learn from a wider range of topics and task types, rather than relying on previously collected data. It also provides accurate control over complexity, ensures semantic consistency, reduces errors, and improves training data quality. WebShaper overcomes the constraints of natural language ambiguity and provides a controllable, explainable, and scalable data synthesis system.
WebShaper-72B was trained with QWEN2.5-72B using data synthesized by WebShaper, achieving cutting-edge results with Gaia Text (60.19). This is a benchmark designed to evaluate common AI assistants in real-world tasks that require users of inference, multimodal understanding, web browling, and traditional tools. We have released a dataset with 500 question answer pairs on Huggingface and ModelScope. Additionally, the smaller model, WebShaper-7B, is open sourced for developers and researchers interested in further research in the field of information exploration.
AMAP's Qwen-driven AI-Native Travel Agent redefines navigation
China's leading digital mapping and navigation platform, AMAP has deeply integrated Alibaba's flagship basic model Qwen, leveraging cutting-edge multimodal and inference capabilities to launch the world's first AI native travel agent. Innovation in spatial intelligence allows users to create complex travel plans through natural voice commands.
AMAP and Tongyi Lab, developers of the QWEN model, collaboratively build a seamless speech interaction system based on Qwen, covering awakening, recognition, understanding and playback. The system also boasts a dual automatic speech recognition (ASR) setup, ensuring high accuracy for both everyday language and interest (POI) information recognition.
At the heart of the system's decision-making capabilities are complex POI inference subagents fine-tuned with the QWEN model. With user outlets, the agent interprets and analyzes multidimensional inputs such as geographic location, child-friendly, participant requirements such as PET allowance, time constraints, transportation preferences, and POI attributes such as seller ratings and opening hours to generate tailored recommendations. For example, questions such as “Find Zjiang cuisine restaurants near Westlake that serves children's meals, with ratings of over 4.5 out of 5, within a kilometer walk from the metro station” could cause accurate suggestions and optimized routes.
The agent integrates its own map API, real-time weather queries, and live traffic tools to provide dynamic guidance such as re-routing to avoid crowds during evening rush hour on the way to the airport.
With over 1 billion users, AMAP is China's top navigation application to date. Meanwhile, Alibaba's Qwen AI model is the world's top open source model with a hugging face platform with over 400 million downloads.
New Agent Coding Platform for Global Developers
Software development AI has emerged as one of the fastest growing areas of the past year, and has evolved beyond simple code completion to become an integral part of the entire development lifecycle. However, in the age of AI, new challenges are revealed and in some cases strengthened. The abstract nature of software complicates knowledge alignment and inheritance, contributing to the friction between humans and cooperation. Furthermore, collaboration with most people remains synchronized and requires a certain period of time, limiting the efficiency of AI.
Against this backdrop, Qoder, an agent coding platform for global developers, has been introduced. Combining advanced context engineering and enhanced knowledge visibility, Qoder is packed with rich features to drive developer productivity.
Featuring its own Next-Edit-Suggestion (NES) model and enhanced context engineering, Qoder is accessible through deep codebase search and queries, multi-line editing, intelligent code suggestions and refactoring, automated testing and verification, all through natural language commands.
Qoder's AI-Native workflow allows developers to streamline the coding process by integrating AI capabilities into their development environment. It is available on Mac and Windows, and supports over 200 programming languages, including Python and Java. Global developers can now experience Qoder with free public previews.
