Chinese agricultural LLMs beyond fisheries — the university-model wave, the SOE entrants, and the honest adoption picture

East-Asia (China); notable concentration in Heilongjiang (Tiangong Kaiwu, Shennong rollout) and Jiangsu (~20 models, per provincial media)

Content

The September 2026 China cycle corrected the corpus’s biggest China-AI conflation — Shennong (CAU) and Sinong (NAU) are two models, not one — and left the rest of the agricultural-LLM population as a named-but-unwritten lead. This unit writes it. The picture that emerges is a university-model wave (CAU, NAU, HIT, Anhui Agricultural University, provincial academies), a smaller SOE wave (Sinochem, China Telecom, China Mobile), and an adoption evidence base that is thin but — unusually for this corpus’s China coverage — includes named farmer-level cases and a Chinese-language critical voice.

The university wave

• Shennong 神农 (China Agricultural University). v4.0 released July 2026 (corpus anchor: pest/disease recognition 615 → 1,016 species at >93%, cau.edu.cn). A 2 September 2026 Jiangsu official-media feature reports 4.0 covering all 31 provinces, “serving a million farmers”, with over 5M mu of applied use — a jump from the >100k users reported at the 3.0 launch (13 October 2025, World AgriTech Innovation Conference) that no source explains. People’s Daily (27 January 2026) supplies the corpus’s first named farmer-level adoption: Du Lianhui, 600 mu of corn near Shenyang, per-mu cost down from ~480 to below 400 CNY in 2025; a 神农卫田 WeChat mini-program for photo-based diagnosis; 36 specialist agents; a “three-sentence” answer design for farmer usability; and the team’s own admission that compute limits prevent unlimited farmer use.
• Sinong 司农 (Nanjing Agricultural University). Released 10 January 2026, open-weight 8B/32B on ModelScope (NAULLM org) and GitHub — the corpus’s only directly verifiable open-weights claim in the Chinese agri-LLM population. Corpus of ~4bn tokens (~9,000 books, 240k+ papers, ~20k policies/standards); multi-agent RAG; explicitly framed as implementing the State Council “AI+” opinion. NAU subsequently released 问梨 (world-first open pear-industry LLM) and 司农智机 (agricultural machinery).
• Tiangong Kaiwu 天工开悟 (Harbin Institute of Technology, Liu Jie team). HIT’s own newsroom (11 March 2026) calls it the first officially registered (备案) agricultural LLM in China and “the first agricultural large model covering the whole agricultural production chain”: growth prediction across 24 crops / 64 varieties; deployed on 114 Heilongjiang farms, serving over 1M mu, with claimed average +5% yield / −10% cost (all HIT-reported). 2024 UK international invention exhibition diamond award; 2025 the only smart-agriculture entry in the China-SCO Digital Technology Toolbox. v3.0 due H2 2026 with an agent platform of 1,200+ agricultural model algorithms. Name discipline: 天工开悟 (kaiwu, “enlightenment”) — not 天工开物, the 17th-century technology treatise the romanisation invites.
• Provincial and sector models. The Jiangsu feature names a dense provincial population: 瑞雪 (global-first plastic-film vertical LLM, Jiangsu Academy of Agricultural Sciences), 豆好 (mung-bean breeding agent with Zhejiang University), 谷语 (first Jiangsu nationally-registered agri vertical model), 新犁 (first Jiangsu model in production decision use), 光伏慧农 (agri-PV trade-off model); Anhui Agricultural University’s 丰菽 2.0 (soybean) and 棃想 (pear); Shanghai’s 丰登 (first bio-breeding model); Heilongjiang’s 龙江农知. Anhui’s Gengyun 耕云 (provincial MARA + iFlytek, first shown September/October 2023, 360M records / ~2M documents) is the cautionary case: a critical Chinese blog reports no findable public entry point, and nothing post-2023 surfaced — existence as a live product rests on one launch announcement.

The SOE wave

• Sinochem iMAP (28 July 2025, Sinochem newsroom): first central-SOE crop-cultivation LLM, built with China Mobile on the 九天 (Jiutian) base model; eight scenarios; ~1M mu of corn pilots in Inner Mongolia and the Northeast; delivered through Sinochem’s several-hundred MAP (Modern Agriculture Platform) service centres — an SOE delivery channel, not an app-store channel.
• China Telecom 雄小农 Xiongxiaonong (Hebei provincial MARA + China Telecom, February 2025): the first DeepSeek-based agricultural sector model, built on China Telecom’s 星辰·神农一号 base — note the naming collision with CAU’s Shennong across two different institutions. Documented at snippet level in a peer-reviewed survey and China Telecom’s ESG report; no Hebei government primary page located; no user numbers anywhere.
• CAAS models: the Institute of Agricultural Information with CNKI released China’s first general agricultural LLM and knowledge platform (August 2024, 1bn+ records); the Institute of Agricultural Resources released 農耕大模型1.0 (December 2025, farmland monitoring, Tang Huajun team); CAU’s 工耕 agri-robotics model (July 2026).

The adoption picture — better than the corpus’s standing finding, still unaudited

The corpus’s September cycle finding was that on-farm usage evidence was almost absent, anchored by 240 million agricultural-LLM tokens/year against ~200 million farm households (units/china-smart-fisheries-fanli-llm.md). This unit upgrades that in one respect and holds it in another:

• Upgraded: named farmer-level adoption now exists (the Du Lianhui case; the 神农卫田 mini-program user), plus institution-reported user counts (>100k Shennong 3.0 users; “million farmers” 4.0 claim; 114 Tiangong Kaiwu farms).
• Held: no independent evaluation of any Chinese agricultural LLM was found — every outcome figure is deployment-actor reported, consistent with the corpus’s standing no-audit finding. The most credible countervailing evidence is internal to the sector: the Jiangsu feature’s own framing that most models “remain in laboratories and demonstration bases”, the named barriers (data silos, cost, elderly-farmer usability, missing mechanisation beneath the AI layer), and CAU’s admission of compute limits on farmer access.

What this unit is doing in the taxonomy

Anchors the Chinese agricultural LLM population as a cell distinct from the fisheries-specific LLM unit and the policy-document chain. Distinct from:

• units/china-smart-fisheries-fanli-llm.md — the fisheries LLM unit where the 240m-token anchor lives; this unit covers the rest of the population and carries the adoption-evidence assessment.
• units/china-smart-agriculture-action-plan-2024-2028.md — the plan layer that mandates “20+ base model algorithms and general software tools or SaaS” by 2028; this unit is the emergent supply those targets describe.
• units/china-no1-central-document-2026-ai.md — the political-mandate layer, refreshed 2026-11 with the 15th FYP “AI + agriculture” action and the intelligent-design-breeding clause.
• units/china-ground-robotics-unmanned-farms.md — where Tiangong Kaiwu also appears, as decision support inside the ground-robotics delivery story.

New gaps surfaced by this unit

• G-476 (new): independent evaluation of Chinese agricultural LLMs — no benchmark, third-party audit or peer-reviewed outcome study of any named model was found; the CSAE survey documents the population without evaluating it. Track: any MRV-style audit, benchmark leaderboard, or academic evaluation of Shennong/Sinong/Tiangong Kaiwu.
• G-477 (new): the Shennong user-count discontinuity — >100k users (October 2025, 3.0) to “million farmers / 5M mu” (September 2026, 4.0) with no intermediate measurement or methodology in any source.
• G-478 (new): 雄小农 primary documentation — no Hebei provincial-government primary page; the model is known at snippet level via the CSAE survey and China Telecom ESG material; no user or deployment figures.

New contested claims surfaced

• C-360 (new): “Chinese agricultural LLMs are reaching farmers at scale” — supported by institution-reported counts (million-farmer claims, 114 farms, 5M mu) and one named farmer case; countered by the sector’s own verdict that most models remain lab-bound (“盆景不是风景”), the absence of any independent evaluation, and the 240m-token/year aggregate that implies per-household usage near zero.