CIDI (03881) has made key progress in exploring AI applications for mining. Will "Xiaoyuan," the world's first mining intelligent agent, be the mining industry's version of Muse?

Title context: CIDI (03881) has made key progress in exploring AI applications for mining. Will "Xiaoyuan," the world's first mining intelligent agent, be the mining industry's version of Muse? Text: Amid the continued leaps in large model capabilities and the rapid decline in computing power costs, AI applications are entering a period of explosive growth. If the release of Meta Muse on the consumer side marks personal Agents entering a new stage from "answering questions" to "getting things done for people," then when AI truly enters the hardest of hardcore industrial sitesmines and portswho can become the intelligent agent that "goes on duty"? CIDI (03881), which has long been deeply engaged in autonomous mining trucks, was the first to give an answer. On September 23, CIDI released the world's first mining intelligent agent, "Xiaoyuan." According to the introduction, "Xiaoyuan" is the industry's first vertical-domain intelligent agent to achieve a full-process closed loop from voice input to task execution in a mining production environment. It debuted in the "Yuan Mine" scheduling middle platform, enabling the management of an entire mine with a single sentence. It is worth mentioning that before its official debut, "Xiaoyuan" had already been put into practical application in 15 mines. Daily operations in mines have long been trapped in an extensive model of "operations by hand, information by staring, decisions by experience." The training cycle for newcomers is long, and on-the-spot responses rely heavily on the personal experience of core dispatchers. This experience-driven model is difficult to standardize and difficult to replicate, forming an invisible ceiling on mine expansion. "Xiaoyuan's" solution is to use a unified intelligent agent entry point to connect the mine's existing digital systems. A dispatcher says one sentence, and the system automatically completes status queries and parameter verification, and after safety checks generates executable instructions. Only operations involving high risk require manual confirmation. This system architecture capable of closed-loop execution consists of three layers. The "Yuanshen" mining large model is responsible for understanding intent, the Harness intelligent agent orchestration engine breaks complex goals into executable task chains, and the heavy-load world model performs physical inference through end-cloud collaboration. Unlike current general-purpose large models on the market that can only answer mining knowledge questions, the "Yuanshen" mining large model is trained on mining corpora and process rules, and uses real-time production data as context to complete intent understanding, task reasoning, and plan generation, allowing rapid adaptation to different mine sites. The difference between the two lies not only in response speed, but also in the depth of understanding of the business and the reliability of the results. This architecture gives Xiaoyuan three underlying core capabilitiesfull-domain monitoring, understanding-based execution, and pre-event prediction. It can maintain 24/7 all-weather duty, achieving full-domain monitoring and scheduling of all mine equipment and operation scenarios; handle faults and unexpected conditions based on on-site status, achieving understanding-based execution; and, through historical data and real-world perception, predict equipment failures, yard congestion, and output trends, moving post-event response forward to pre-event intervention and directly raising the ceiling for capacity and safety. Real-world data confirms this. From the operational data sent back by the 15 mines where "Xiaoyuan" has been deployed, the three core indicators of safety, efficiency, and cost have all improved significantly. On the safety level, safety reviews passed on the first attempt, scheduling instructions were fully traceable throughout the process, data was retained on the mine's intranet, and safety management specifications required zero changes. On the efficiency level, batch scheduling was compressed from several minutes to 3 seconds, production reports were shortened from most of a day to automatic generation in 3 minutes, and fault handling efficiency increased by 80%. On the cost level, unit transportation cost decreased by 5%, single-vehicle transportation efficiency increased by an average of 11.7%, and 100 autonomous mining trucks transported an additional 10,000 cubic meters of material per day. As a new exploration by CIDI in mining AI applications, "Xiaoyuan" has significantly improved productivity. It has turned mine scheduling from something highly dependent on personal experience into a replicable system capability. From the perspective of the entire industry's development, the entry of large models into mines also means that mining's exploration of AI has entered deep water. Following this logic, as "Xiaoyuan" is deployed at scale in more mines in the future, this "replicable system capability" is expected to accelerate its penetration across the entire industry and push the industry's level of intelligence to a new level.