China Agricultural University Deploys "Shennong" Agricultural LLM (神农大模型) to 100,000+ Chinese Farmers — 36 Specialized Agents, 3 PB Knowledge Base

Title

China Agricultural University Deploys "Shennong" Agricultural LLM (神农大模型) to 100,000+ Chinese Farmers — 36 Specialized Agents, 3 PB Knowledge Base

Content

China Agricultural University (中国农业大学) built and operates the "Shennong" (神农) agricultural large language model — a vertical-domain AI for agriculture that, in version 3.0 released October 2025, deploys 36 specialized agents and serves more than 100,000 farmers across China. The model covers 90% of agricultural disciplines and 80% of agricultural scenarios, with a core knowledge base of 10 million agricultural knowledge graph entries, 20 million labeled images, and 50 million production records (3 PB total). Per Wang Yaojun (王耀君), Associate Professor at CAU's School of Information and Electrical Engineering, who leads the Shennong project: "Agriculture is an industry employing nearly 200 million people. We must use AI to drive the shift from 'depending on the weather' to 'knowing and acting with the sky.'" The motivation: Wang and team observed in 2020 that general-purpose LLMs often gave "layman answers" or even常识-violating responses on agricultural questions because crops, soils, and climates vary drastically by region. Building a vertical domain model required proprietary data, not just scraping public agricultural documents. The Shennong team scanned 3,000+ books in the CAU library over 7 months (combined with public material, building a 20,000-book specialized database), then traveled to 20+ provinces collecting real soil composition, irrigation records, pest records, and meteorological data. To cope with limited GPU budget, the team adopted a Mixture-of-Experts (MOE) architecture combined with model compression and pruning, cutting compute cost while keeping performance. Version 1.0 (December 2023) covered Q&A, semantic understanding, summarization, and decision reasoning; Version 2.0 (July 2024) added multimodal image and audio recognition; Version 3.0 (October 2025) adopted "lightweight + multi-agent" architecture — halving compute consumption while launching 36 specialized agents for specific farming tasks. Operational scale: 3.0 deployed since October 2025, serving 100,000+ farmers; 36 specialized agents; 90% ag disciplines covered; 80% ag scenarios covered; 10M knowledge graph entries; 20M labeled images; 50M production records; 3 PB total. Real outcomes: Shenyang farmer Du Lianhui cut corn management cost from ~480 yuan/mu (2024) to <400 yuan/mu (2025) on 600 mu using the system; Beijing Huairou "Yanqi No. 2" lettuce bred by Shennong's smart-breeding agent; "policy navigator" agent piloted in Heilongjiang townships to lighten the burden of frontline policy communication. Per Wang: "The emergence of vertical agricultural large models will significantly lower the entrepreneurship threshold and technical barriers for 'new farmers,' enabling more young people aspiring to agriculture to master core agronomy through AI tools." The team plans continued iteration to push agriculture from "experience-driven" to "data-driven." Sources: 央广网 (CNR), 2026-01-28, by 央广网 reporter. Note: originally published in Chinese; this English narrative was translated for cross-language discoverability. Original Chinese text preserved in raw archive.