Midnight on a fish farm used to mean a flashlight, a tired farmer, and a guess about whether the water was turning bad. Today it can mean a dashboard that already adjusted the feeding schedule two hours ago and texted the owner that dissolved oxygen is fine.
Chinese aquaculture is enormous and famously low-tech. That contrast is exactly why it is becoming an AI story. The model doing the work has a name: 北冥智渔 (Beiming Zhiyu, "Northern Sea Fishery Model").
The model built from 2,000 farms
Litugaoke (励图高科), a Shandong marine-technology company, built 北冥智渔 on 72 billion parameters, fusing real-time data from more than 2,000 aquaculture enterprises with the accumulated experience of 100,000 farmers. The result is described as a "water-feed-disease" (水—饵—病) AI decision network: it watches water quality, recommends feeding, and flags disease risk before a pond turns.
On the ground, at the Weishan Lake (微山湖) modern fishery industrial park, the model became a "smart fishery digital-management platform" covering both pond and recirculating-aquaculture modes. The operator quoted by Shandong's development commission describes the old routine — up before dawn to patrol, judging oxygen shortage by watching fish at the surface, feeding by rough estimate — replaced by a system that adjusts feed by how eagerly the fish eat and how much is left. The cited results: labor down about 70%, feed down about 15%, and roughly 1 million yuan (≈ US$141,000 / HK$1.1 million) saved per year.
Why fish are harder than they look
Aquaculture is a brutal environment for AI. Underwater imagery is murky; conditions shift by the hour; a single misread sensor can wipe a harvest. Litugaoke's engineers hit a "data famine" at the start, buying data from farms and later partnering with the National Fisheries Technology Extension Center to open the spigot. They then attacked the hard parts: a blockchain-based fishery knowledge graph to tidy messy data, a BERT-style model to extract key signals, a vision transformer (ViT) for fine-grained image features, and an improved deep-Q-network as a smart task router that learns the best way to distribute compute.
The payoff is a closed loop — data from real farms trains the model, the model improves those farms, and the better farms generate cleaner data. In 2025 the model was named among Shandong's first batch of typical large-model application scenarios and a Qingdao smart-ocean benchmark.
Shrimp by the cubic meter
The broader surge shows up across Guangxi, a provincial leader in smart fisheries. At an automated shrimp workshop in Dongxing, a visual neural network tracks shrimp growth; yield reaches 15 kg per cubic meter. Guangxi Huaxing's AI facility-fishery system pushes California bass to 50 kg per cubic meter. In 2024 Guangxi's aquatic output was 3.887 million tons (up 2.7%) and its fishery economy passed 127.6 billion yuan (up 2%) — the kind of gain precision farming is meant to deliver.
The pattern repeats: AI here does not replace the farmer, it removes the 3 a.m. panic. As one Guangxi farmer put it, the phone now shows water quality and controls feeding, and "one person does what ten used to."
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, aquaculture is AI's next farmland — "the data is messier than the office, but the margin it protects is someone's whole year of income."
Honest limitations
Core figures come from Shandong's provincial development commission, Economic Daily/China Economic Net, and People's Daily Guangxi, quoting the company and farmers; they are company- and state-media disclosed, not independently audited. The "72 billion parameters" and the 70%/15%/1-million-yuan savings are vendor-stated outcomes from specific pilot sites, not a statistically sampled average across China's millions of ponds. Underwater sensor reliability and connectivity in remote areas remain real bottlenecks, and smallholders without the digital base may not see these gains soon.
What readers can do now
- If you invest in agri-tech, watch fishery and livestock large models — they turn thin-margin, data-poor farming into a measurable-efficiency game.
- If you farm or advise farmers, start with one sensor loop (oxygen + temperature + auto-feeder) before any model; clean data is the precondition, not the model.
- If you follow food security, note that precision aquaculture raises yield per cubic meter — relevant as land and water pressure tightens on traditional protein sources.