A farmer in a village snaps a photo of a spotted leaf. Before the agronomist calls back, an app has already named the disease, suggested a treatment, and flagged which neighboring plots are likely next. The expert still decides — but the wait, and the guesswork, are gone.
That is the promise behind Shennong agricultural large model (神农大模型) 4.0, released by a team at China Agricultural University (中国农业大学). It is framed less as a product launch than as a shift from "can it recognize the bug?" to "can it reason about the whole field?"
Key takeaways
- Shennong 4.0 (神农大模型 4.0), from China Agricultural University (中国农业大学), recognizes 1,016 crop pest and disease types at over 93% accuracy.
- A companion Shennong World Model (神农·农业世界模型, SNWM) 1.0 was released and open-sourced, letting users simulate a field digitally before acting in the real one.
- The world model was validated on tomato, lettuce, and cucumber crops and ran a drought scenario across 38 counties in Jilin province.
- By the platform's own accounting it already reaches 31 provinces, over 5 million mu (roughly 330,000 hectares) of farmland, and more than 1 million farmers.
From recognizing to reasoning
The headline number — 1,016 crop pest and disease types at 93% — matters, but the more interesting shift is architectural. Traditional agricultural AI is good at identification: show it a leaf, it names the blight. Shennong's team, led by associate professor Wang Yaojun (王耀君) at the university's College of Information and Electrical Engineering, pairs the recognition model with a world model that holds a causal, simulatable picture of a farm.
Think of it as a digital twin that grows. Instead of only answering "what is wrong with this plant?", the world model can answer "if I irrigate three days early and raise the greenhouse temperature by two degrees, how will the crop evolve, and what happens to yield and risk?" That is a counterfactual question — the kind a seasoned agronomist answers from memory and intuition.
SNWM 1.0 was released on 12 July 2026 and open-sourced, according to China Agricultural University's news office and People's Daily. It was tested on tomato, lettuce, and cucumber in controlled settings, and ran an action-conditioned drought simulation across 38 counties in Jilin province, where the model projected how different management choices would shift yield under dry conditions.
How the model got here
Shennong is not a one-shot. Version 1.0 (2023) handled agricultural knowledge Q&A. Version 2.0 (2024) added multimodal input — text plus images. Version 3.0 (2025) broke down disciplinary walls to cover about 90% of agricultural subjects and 80% of scenarios, and could run offline for data security. Version 4.0 (July 2026) extends perception, adds genomic analysis for breeding, and opens standardized API and MCP interfaces so third-party systems and agents can call its recognition, Q&A, and reasoning.
One practical bridge to real users is language. The team released Shennong AgriTalk, which supports several Chinese dialects, so a farmer in a rural county can speak to the model in the local tongue rather than standard Mandarin. That detail matters more than it sounds: agricultural AI fails at the last mile if the interface assumes a literate, Mandarin-speaking user with a quiet connection.
Scale and the "buy service, not hardware" gap
The platform reports coverage of 31 provinces and over 5 million mu (roughly 330,000 hectares) of farmland, reaching more than a million farmers. Those are the company/university-stated deployment figures, not an independently audited census, but they signal a move from pilot plots to regional rollout.
The harder problem in Chinese agriculture is not the model — it is the fragmented, aging, small-plot reality. A separate, nearby example shows the delivery model that makes such models usable: in Dongguan's lychee orchards, the local government built a "sky-ground" sensing network (drones, growth monitors, AI pest traps, spore catchers) feeding a provincial smart-agriculture brain, and then sold the service to smallholders rather than making them buy equipment. Inspection efficiency there was reported as substantially higher, with pest and disease recognition accuracy over 90%. The lesson travels: a 93%-accurate model only helps if someone delivers it to a smallholder working a small plot.
What readers can do now
- If you farm or advise farmers, treat pest-ID accuracy above 90% as table stakes, not a differentiator. The real question is whether the tool also reasons about timing, dosage, and neighboring plots.
- Ask how the model reaches the field. Open APIs and dialect support matter more for adoption than raw species count.
- Pilot the "world model" mindset: before a costly input decision, ask what simulation or historical counterfactual the tool can show you — not just what it can label.
Honest limitations
The 1,016 pest-and-disease type count, 93% accuracy, the July 2026 release dates, the open-sourcing of SNWM 1.0, the tomato/lettuce/cucumber validation, and the Jilin 38-county drought test all come from China Agricultural University's news office, People's Daily (via CAU), and the Xiaoshan District government — mutually consistent authoritative sources. The "31 provinces / 5 million mu / 1 million farmers" reach is stated by the Xiaoshan government write-up of a Shennong roadshow and should be read as the platform's own deployment claim, not an independent audit. I did not independently verify the 93% figure against a held-out test set, and accuracy in the field will vary with image quality and crop mix. The Dongguan lychee example is drawn from a Guangdong Agriculture Department account and is included only as an illustration of the delivery model, not as part of the Shennong system. I have not covered livestock, aquaculture, or smart-greenhouse AI in depth.