ILUVATAR COREX(09903) 2025 performance report released: Revenue from the reasoning series increased by 238.2% year-on-year, with both training and competition tracks making strong efforts.

On March 30, ILUVATAR COREX (09903) released its first performance report after its listing on the Hong Kong Stock Exchange. The company achieved annual revenue of 1.034 billion yuan, a year-on-year increase of 91.6%; gross profit of 558 million yuan, a year-on-year increase of 110.5%, with the gross profit growth rate exceeding the revenue growth rate, and the product profitability steadily improving. Adjusted net loss narrowed by 32.1% year-on-year, research and development efficiency improved in sync with operational efficiency, the financial structure remained stable, and performance achieved high-quality growth. The core general-purpose GPU business has become the main pillar of revenue, with annual revenue of 923 million yuan, a year-on-year increase of 149.6%, accounting for 89.3% of total revenue. Among them, the Tianyi training series revenue was 584 million yuan, a year-on-year increase of 116.7%, and the Zhikai reasoning series revenue was 339 million yuan, a year-on-year increase of 238.2%. Both training and reasoning tracks achieved rapid breakthroughs, fully adapting to market demand. Leveraging existing general-purpose GPU technology accumulation, the company has launched the Tongyang series end-side computing products targeting Siasun Robot & Automation, intelligent terminals, and other fields, to comprehensively layout cloud-edge-end products. By the end of 2025, ILUVATAR COREX has served more than 340 customers in various industries, deploying products and solutions in over 1,000 projects in fields such as the internet, AI large models, research, finance, medical, education, and transportation, continuously expanding its commercial scale. The company's technological research and development have also made continuous breakthroughs in addressing large model reasoning scenarios. Through continuous investment in core technologies such as PD separation (decoupling of Pre-fill and Decoding stages), lossless quantization, and cumulative parallelism, they have improved resource utilization and business SLA indicators for large model reasoning. By providing a large model acceleration toolkit with lossless quantization technology, they have achieved a more than 60% increase in computing power utilization while fully guaranteeing model reasoning accuracy. With support from the DeepSpark open-source community, they have adapted to over 610 mainstream algorithm models and possess Day0 native support capabilities for new models.