The "Tong-Yun-Ge" golden triangle has emerged. On the morning of January 29, T-Head's official website quietly launched a high-end AI chip named "Zhenwu 810E," which is the previously leaked Alibaba self-developed PPU. This marks the first official appearance of Alibaba's AI golden triangle, "Tong-Yun-Ge," composed of Tongyi Lab, Alibaba Cloud, and T-Head. The "Zhenwu" PPU has already been deployed in multiple 10,000-card clusters on Alibaba Cloud, serving over 400 clients including State Grid, Chinese Academy of Sciences, XPeng Motors, and Sina Weibo.
According to T-Head's official website, the "Zhenwu" PPU adopts a self-developed parallel computing architecture and inter-chip interconnect technology, paired with a full-stack self-developed software stack, achieving complete hardware and software self-sufficiency. It features 96G of HBM2e memory and an inter-chip interconnect bandwidth of 700 GB/s, making it applicable for AI training, AI inference, and autonomous driving. Alibaba has already utilized the "Zhenwu" PPU extensively for the training and inference of the Qwen large model, deeply optimizing it with Alibaba Cloud's comprehensive AI software stack to provide customers with integrated products and services.
Industry insiders reveal that, based on key parameter comparisons, the overall performance of the "Zhenwu" PPU surpasses NVIDIA's A800 and mainstream domestic GPUs, and is comparable to NVIDIA's H20. Industry practitioners note that the "Zhenwu" PPU boasts excellent and stable performance with outstanding cost-effectiveness, earning a strong reputation within the industry where demand currently outstrips supply. The official launch of the "Zhenwu" PPU showcases the substantial capabilities T-Head has accumulated over years in the chip field.
Looking back, Alibaba founded Alibaba Cloud in 2009, established the T-Head chip company in 2018, and initiated large model research in 2019. After 17 years of strategic investment and vertical integration, the company has finally realized the complete layout of the full-stack AI "Tong-Yun-Ge." On January 26, Tongyi Lab released the Qwen flagship inference model Qwen3-Max-Thinking, which set new global records in multiple authoritative evaluations, with performance comparable to GPT-5.2 and Gemini 3 Pro.
This model has a total parameter count exceeding one trillion (1T) and was pre-trained on a massive dataset of 36T Tokens, making it currently Alibaba's largest and most powerful Qwen inference model. Across 19 recognized large model benchmarks covering factual knowledge, complex reasoning, instruction following, human preference alignment, and Agent capabilities, the Qwen flagship inference model broke several state-of-the-art (SOTA) records, with overall performance rivaling 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 use, the Alibaba Tongyi team further trained the model on a wide variety of tasks using a combination of rule-based rewards and model-based rewards in a joint reinforcement learning process, 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 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. According to the latest data from the world's largest AI open-source community, Hugging Face, the number of derivative models based on the open-source Qwen model has surpassed 200,000, with downloads exceeding 1 billion, firmly securing its position as the global leader.