In 2026 year 9 month, at the Eighth China Swine Industry Science and Technology Conference, Associate Professor Li Xiang of China Agricultural University delivered a report titled “The Current State of Artificial Intelligence + Livestock Farming and Further Reflections.” Associate Professor Li Xiang reviewed the implementation of AI farming technologies over the past ten years, introduced the team’s self-developed agricultural vertical large language model “孺子牛,” and shared the futurePig farm constructionApproach.
01 AI aquaculture technology implementation over the past ten years
The first part of the report reviews the implementation of AI technologies in animal farming over the past ten years and groups them into three areas: precision sensing, precision decision-making, and precision management. Precision sensing relies on contact and non-contact methods, with vision and sound perception at its core. Using deep learning, it enables individual livestock and poultry identification and behavior recognition, and collects multidimensional information such as body weight, body condition, activity, feed intake, limb condition, and environmental parameters. Applications include monitoring pig body measurements, lameness, and aggressive behavior, as well as intelligent poultry inspections and detecting abnormal nighttime behavior in pigs. Precision decision-making is exemplified by intelligent disease diagnosis. It uses thermal infrared and visible-light images to identify and provide early warnings for livestock and poultry diseases, estrus, and feeding and rumination behavior; some detection items have already achieved relatively high accuracy. Precision management is implemented through farm ERP management systems covering environmental control monitoring, production processes, and batch management, complemented by precision feeding, environmental regulation, and intelligent manure and wastewater treatment to connect production and environmental protection in a closed loop. Overall, AI in animal farming has moved beyond isolated pilot projects and developed into a systematic set of practical applications.
02 Agricultural vertical large model “孺子牛”
Part two introduces the team’s independently developed agricultural vertical large model, “孺子牛.” The model has accumulated 9 years of big data from the dairy cow industry, covering hundreds of farms across multiple provinces. It encompasses structured data such as production and health data, as well as unstructured data such as video images and management reports. Based on GLM-4-9B, the model was fine-tuned using a domain knowledge base and expert knowledge to build a super brain for farms, a decision-making system, and digital twin farms. Compared with general-purpose large models, “孺子牛” provides clear metrics and more practical answers to industry questions, including specific feed ratios and additive dosages. In practical applications, the model can predict the body weight of replacement heifers and the milk production of first-lactation cows, helping farms increase milk production. It can also use DHI data analysis to identify production problems on farms and optimize grouping and culling plans, thereby increasing farm profits and reducing nitrogen loss. The report states that the upper limit of AI in a vertical industry is determined by the extent of the industry’s data and accumulated professional knowledge.
03 Future pig farm construction concepts
Part three returns to AI pig farming and explores approaches to building future pig farms. Existing industrialized pig farms have achieved standardized production, but can only be managed based on average indicators, overlooking individual differences among pigs and making it difficult to achieve dynamically optimal production. Pigs may be the same age, but their genetics, health, and environments differ. In the future, AI will need to combine individual, historical, and environmental data to provide differentiated farming strategies.
Achieving this goal rests on three key technical pillars: first, building a high-quality dataset that connects the individual, the environment, and the production chain, providing a foundation for intelligent systems; second, developing a world model (digital twin) that goes beyond simply recording the current state to simulate how the herd may develop under different actions, upgrading basic anomaly monitoring into a production decision-making tool; and third, a Harness orchestration engine that integrates multiple models with different functions, breaks down tasks, combines results, and synthesizes multidimensional information to produce actionable production plans, while also establishing a follow-up review mechanism. Agricultural expertise must also be converted into a knowledge graph that AI can recognize, covering the full scope of pig farming, including breeding, farrowing, fattening, and operations. This graph should define the data accumulation, model predictions, and AI decision-making objectives for each stage, ultimately creating an AI pig farm capable of thinking and evolving independently.