From custom projects to a general asset: VisionAgent debuts in Shanghai

On 26 March, at the Vision China 2026 Shanghai forum, Beijing Lingbang Intelligent Equipment launched VisionAgent 3.0, a new general platform for industrial vision.

The booth stayed crowded. Engineers and technical leaders from many manufacturing sectors tried the system and asked how to bring it onto a line and deploy it as a platform.

The questions on the stand were consistent: plants want to leave one-off custom vision projects and buy a general visual product. From 1.0 in October 2024, to 2.0 in 2025, to 3.0 now, Lingbang has been iterating against that demand.

In the keynote, chairman Cui Zhongwei walked through live cases and the architecture and commercial logic of VisionAgent 3.0.

Commercial logic: from a custom project to a general asset

Traditional machine vision often means one algorithm per machine—and a dead system after a line change. Cui put it this way: most hardware on a plant floor is a dedicated asset; a general asset is what shows reuse value on the books. VisionAgent 3.0 is meant to be that general asset for industrial vision.

The 3.0 rebuild sits on five engineering axes.

1. Model base: 1.3B parameters, few-shot start

Classic vision projects train a small model on a huge pile of defect photos for one scene. VisionAgent 3.0 carries a foundation model Lingbang trained over 15 months from January 2023—100 million real industrial images, 1.3 billion parameters. On a live line, process staff can set a high-accuracy task with 1–5 sample images. That cuts the time spent collecting and labeling data.

2. Compute: RTX 4090, 250 FPS

To keep a large model responsive at the edge, VisionAgent 3.0 runs on an RTX 4090 and has measured 250 frames per second—millisecond-class detection and feedback on a high-takt automated line.

3. Hardware: all-in-one, lower TCO

One agent is the node: up to 16 industrial cameras, 31 lights, 12 actuators and 6-axis motion. This “one brain, many eyes” layout skips a pile of industrial PCs, PLCs and light controllers, shrinks the cabinet, and cuts total cost of ownership.

4. Interaction: zero-code drag-and-drop, plus a state machine

Industry needs reliability more than a chatbot. The UI is fully visual: no code to write a detection flow. For actions a single frame cannot judge, a state machine reads multi-frame sequences—such as an SOP—and applies logic across time.

5. Delivery: days, not months

Decoupled software and hardware show up as a shorter delivery cycle—the number executives asked about most on the stand. VisionAgent 3.0 compresses month-scale custom development into a few days of standard product rollout, which pulls ROI forward.

Beijing Lingbang Intelligent Equipment was founded in 2016. In its tenth year, VisionAgent 3.0 marks a full product-strategy step: industrial vision moving from custom projects toward a standard, general era.