Annual Report 2025

INVESTMENT MANAGER’S DISCUSSION AND ANALYSIS (CONTINUED) China Merchants China Direct Investments Limited Annual Report 2025 29 REVIEW OF INVESTMENTS (CONTINUED) In 2025, Hanwei continued to advance the downstream applications of its second-generation UWB product, MK8080 chip. The Industrial Internet of Things (IoT) was further promoted and has begun downhole application deployment for key customers. In terms of automotive applications, the company has conducted further cooperation with first- tier companies to develop and explore the application of UWB in car keys and radar. In terms of consumer IoT applications, Hanwei has tried to promote collaboration in the products’ research and development. Moonshot AI Ltd. (“Moonshot AI”) was established in Cayman Islands in May 2023. It is a startup company that is focused on the research and development of general-purpose large multimodal model and end-customer applications. The Fund invested US$10 million in December 2023 for a 1.29% equity interest in Moonshot AI. Moonshot AI completed four new rounds of equity financing in February, March and June 2024, and March 2025, respectively, and the equity interest held by the Fund in Moonshot AI was diluted from 1.29% to 0.66%, accordingly. During January and February 2026, Moonshot AI once again completed a new round of financing, as a result of which the Fund’s equity interest in Moonshot AI was diluted from 0.66% to 0.58%. As of the end of 2025, the carrying value of the Fund’s interest in Moonshot AI was US$16.76 million, representing an increase of 9.19% over US$15.35 million at the end of last year. In 2025, Moonshot AI continued to advance its self-developed models. In January 2025, Moonshot AI released the Multimodal Thinking Model, K1.5, which has surpassed certain global benchmark models on test sets covering mathematics, coding, vision-based multimodal tasks, and general ability. In February 2025, Moonshot AI introduced its Mixture of Block Attention (MoBA) and its technical report, which further optimised the model’s long-context understanding ability and its overall performance. In April 2025, Moonshot AI successively open-sourced the Kimi- VL and Kimina-Prover models, both of which demonstrated leading performance in terms of basic function, long context, and multimodal reasoning capabilities. The open-source initiative is expected to enhance the company’s technological influence. In June 2025, Moonshot AI launched Kimi-Researcher, a foundational pre-trained model, which it tested in small-scale. Kimi-Researcher is an end-to-end reinforcement learning-trained agent model capable of generating comprehensive, in-depth research reports through integrated retrieval and autonomous reasoning. In July 2025, Kimi released the open-source K2 model, with a total of 1 trillion parameters supporting 128K long-context input, achieving leading performance across coding generation, mathematical reasoning, and tool utilisation benchmarks, and ranked 11th globally on the LMArena leaderboard (October 2025). In November 2025, Kimi released the open-source reasoning model K2-Thinking, which achieved State-of-the-Art (SOTA) performance across multiple benchmarks, including Humanity’s Last Exam, BrowseComp (autonomous web browsing capability), and SEAL-0 (complex information gathering and reasoning). The model also demonstrated comprehensive improvements in agentic search, agentic programming, writing, and general reasoning abilities.