Alibaba has officially released its flagship reasoning model, Qwen3-Max-Thinking. It is understood that through the scaling of total parameters, reinforcement learning, and computational inference, the new Qwen model has set new global records on several key performance benchmarks, including scientific knowledge (GPQA Diamond), mathematical reasoning (IMO-AnswerBench), and code programming (LiveCodeBench).

The strongest Qwen model was unveiled on the morning of January 27, according to a message from Alibaba Cloud's official WeChat platform. This model boasts a total parameter count exceeding one trillion (1T) and was pre-trained on a massive dataset of 36T Tokens, making it Alibaba's largest and most capable Qwen reasoning model to date.

Across 19 recognized large model benchmark tests covering factual knowledge, complex reasoning, instruction following, human preference alignment, and Agent capabilities, the Qwen flagship reasoning model has broken several state-of-the-art (SOTA) records, with overall performance comparable to GPT-5.2-Thinking-xhigh, Claude Opus 4.5, and Gemini 3 Pro.

Simultaneously, gearing up for the impending era of intelligent Agents, Qwen3-Max-Thinking has also enhanced its native Agent capabilities for autonomously calling tools. Specifically, after initial fine-tuning for tool usage, the Alibaba Tongyi team further trained the model on a large number of diverse tasks using a combination of rule-based rewards and model-based rewards in a reinforcement learning framework, endowing Qwen3-Max-Thinking with a smarter ability to think in conjunction with tools.

This adaptive tool-calling capability can be experienced on QwenChat, where the model autonomously selects from three core Agent tool functions: search, personalized memory, and a code interpreter, while also exhibiting reduced model hallucination. Currently, developers can freely experience the Qwen3-Max-Thinking model on QwenChat, enterprises can access the new model's API services through Alibaba Cloud's Bailian platform, and general users can also try the model via the Qwen PC client and web version.

It is understood that the Qwen APP will also soon integrate the new model, making it accessible to all users. The Qwen APP has already been integrated into the Alibaba ecosystem. Earlier, on January 15, Alibaba officially announced that the Qwen APP has been fully integrated with Alibaba ecosystem services such as Taobao, Alipay, Taobao Flash Sales, Fliggy, and Amap, enabling AI-powered shopping functions like ordering food, shopping, and booking flight tickets, and has been opened for testing to all users.

Concurrently, the Qwen APP's Task Assistant has begun invitation-only testing. Alibaba Group Vice President Wu Jia revealed that this upgrade will launch over 400 AI service functions. "After acquiring a super-powered brain, AI is now growing hands and feet that can reach the real world, practically 'working' for users in their daily lives," Wu Jia stated, adding that the era of AI services is just beginning, with some capabilities still under exploration.

Tests show that after downloading the latest version of the Qwen APP, users can select "My" - "Application Authorization Management" in the lower left corner of the app to authorize access to Taobao, Taobao Flash Sales, Fliggy, Alipay, Alipay AI Pay, etc., and invoke the corresponding permissions. Wu Jia indicated that for consumer scenarios, where marketing information online is complex and noisy, training the model's understanding and discrimination capabilities is crucial. The Qwen APP, while relying on world knowledge, can also leverage Alibaba's transaction and service data to enhance the model, thereby ensuring the objectivity and accuracy of the AI shopping functions.

At the life services level, the Qwen APP has integrated Alipay's government services and achieved cross-application collaboration capabilities, for example, invoking Fliggy's capabilities to complete flight and hotel bookings, and using Amap for itinerary planning. It is understood that, based on the underlying technical capabilities of Qwen, the "Task Assistant" function has already initiated directed invitation testing on both the APP and Web platforms. This function possesses human-like multi-step planning capabilities, covering core scenarios such as application development, Office productivity, consulting research, and daily life services. After the testing phase concludes, this function will be made available to users free of charge.

Earlier, informed sources had revealed that within the coming months, Alibaba would gradually add intelligent AI (agentic-AI) functions to the Qwen application to support shopping features on platforms including the main Taobao marketplace. It now appears that the aforementioned process is being progressively realized. "Alibaba plans to eventually expand globally through overseas versions," an informed source stated earlier, adding that over the past few months, Alibaba CEO Eddie Wu Yongming has reassigned over a hundred developers from various departments to work on this project.

This is also part of the additional AI infrastructure investment announced by Alibaba for 2025. Eddie Wu had previously outlined his plans to launch new models and "full-stack" AI technology, reflecting Alibaba's intent to develop both the services and the underlying infrastructure supporting the technology. Moving from AGI to ASI, Eddie Wu stated at last year's Apsara Conference that large models are the next-generation operating system, and AI cloud is the next-generation computer. He suggested that perhaps only five or six super cloud computing platforms will exist worldwide in the future.

Alibaba is now actively advancing a 380 billion yuan AI infrastructure construction plan and intends to commit even larger investments. Wu Yongming believes that achieving AGI (Artificial General Intelligence) is a certainty, but it is merely a starting point; the ultimate goal is to develop ASI (Artificial Superintelligence) capable of self-iteration and comprehensively surpassing humans to address major scientific challenges like climate change, energy, and interstellar travel.

The path to superintelligence is divided into three stages: first, "emergence of intelligence," where AI acquires generalized intelligence by learning human knowledge; second, "autonomous action," where AI masters tool use and programming capabilities to "assist humans," which is the industry's current stage; and third, "self-iteration," where AI achieves autonomous learning by connecting to the full spectrum of raw data from the physical world, ultimately enabling it to "surpass humans."

Wu Yongming also mentioned at the time that in the great transformation from AGI to ASI, large models will be the next-generation operating system. "This doesn't mean large models will replace operating systems like Windows or Linux. Rather, large models and related systems will assume the role of current operating systems in the interaction between the physical and digital worlds. In the future, almost all tool interfaces linking the real world will connect with large models, and all user demands and industry applications will execute tasks through tools related to large models."