Key Takeaways
- Nanjing Agricultural University released the Sinong language model at a national higher education forum.
- The model is China’s first open-source vertical large language model focused on general agriculture.
- Sinong is trained on more than 4 billion tokens of specialized agricultural data.
- The model is available in 8B and 32B parameter versions via open-source platforms.
- The release supports national strategies on agricultural modernization and artificial intelligence.
Nanjing Agricultural University Launches Sinong Agricultural Language Model
Nanjing Agricultural University officially released the Sinong language model on January 10 during the sub-forum titled “Digital Technology Reshaping the All-Dimensional Transformation of Agricultural and Forestry Education” at the 2025 Annual Meeting of the Higher Agricultural and Forestry Education Branch of the China Association of Higher Education.
According to the university, Sinong is the first open-source vertical large-scale language model in China designed for the general agricultural field, and the first agricultural large language model led by Nanjing Agricultural University. The launch represents a new development in the university’s research and application of foundational artificial intelligence models tailored to agriculture.
The name “Sinong” (司农) originates from an ancient Chinese official title associated with agricultural and financial management, reflecting the model’s sector-specific focus.
Nanjing Agricultural University Builds Domain-Specific Data Foundation
The Sinong language model is built on a large-scale agricultural knowledge base developed using Nanjing Agricultural University’s disciplinary strengths. The research team compiled data from multiple agricultural subfields, including animal science, crop breeding, horticulture, plant protection, veterinary medicine, agricultural economics, smart agriculture, and agricultural resources and environment.
In total, the dataset includes more than 4 billion tokens derived from nearly 9,000 books, over 240,000 academic papers, close to 20,000 policy and standards documents, and extensive online knowledge sources. The university stated that this process resulted in a relatively complete and high-quality agricultural foundation dataset.
“The model focuses on serving the agricultural sector and building a reliable knowledge base for research, education, and production,” the university said in its release.
Technical Design Addresses Accuracy and Knowledge Challenges
To address common challenges such as hallucinations and knowledge lag in professional large language models, the Sinong development team implemented multiple technical approaches. These included instruction fine-tuning combined with thought-chain data and contextual references to improve reasoning and content generation in agricultural contexts.
A multi-agent retrieval enhancement framework was also introduced to support efficient use of domain literature. The framework integrates optimized knowledge base construction, intelligent query rewriting, and hybrid retrieval strategies to improve accuracy and timeliness.
Open-Source Strategy and Future Development
The Sinong language model has been fully open-sourced on the Magic Tower community and GitHub, with 8B and 32B parameter versions available. Nanjing Agricultural University stated that the open-source approach is intended to lower barriers for agricultural AI development and encourage collaboration across research institutions, enterprises, and developers.
The university noted that the release aligns with national policies on building a strong agricultural sector and advancing the “Artificial Intelligence+” action plan. Future work will focus on model iteration, expanded application scenarios, and supporting the digital transformation of agriculture in China.
