Generative AI is accelerating from "chat-only" question-and-answer tools to "super digital employees" that can roll up their sleeves and get to work. On June 8, CloudWalk (9678.HK) officially launched its new general large language model, U2. As a native agent large model tailored for individuals, developers, and enterprise organizations, U2 has completely moved away from the limitations of traditional models that heavily rely on single-turn dialogue. Instead, it firmly anchors its technical focus on delivering high intelligent density and high Token value.

In practical business scenarios, traditional models often struggle with complex system engineering tasks, as they are only capable of handling short-chain text generation. However, U2 has significantly enhanced its continuous execution loop for real-world tasks. In high-difficulty scenarios such as complex office work, software engineering, deep research, and multi-tool collaboration, it can autonomously break down macro-level abstract tasks and continuously advance complex workflows of over 100 steps, achieving a transition from "passive response" to "active execution."

With the release of the new model, the latest results from a series of authoritative capability evaluations have also been released. U2 has reached the first tier of mainstream large models in multiple key dimensions. In the GPQA Diamond evaluation, which tests hard-core knowledge and complex logical reasoning capabilities, U2 scored an impressive 87.9, surpassing many strong competitors in the industry, including GLM-5.1, Hy3preview, DeepSeek-V4-Flash (High), and MiniMax M2.7. This result demonstrates its strong stability when dealing with high-difficulty professional questions.

Aside from excelling in logical reasoning, U2 also delivers impressive performance in daily white-collar work. In the GDPval evaluation, which tests real-world office and knowledge work delivery capabilities, the model achieved an excellent score of 72.5. Unlike conventional rote memorization-based evaluations, GDPval focuses on testing a model's practical output in enterprise production environments. The test results show that U2