Book a consultation

Cognex Deep Learning and AI: Edge AI, OneVision and VisionPro Deep Learning

Cognex OneVision: The Cloud AI Platform and the Questions to Ask

Cognex OneVision is a cloud platform for building, validating and deploying AI inspection models: what it does, what is not public, and IT/OT questions.

Vladimir Romanov


Cognex OneVision is a cloud platform for managing the life of an AI inspection model: developing and labelling it, validating it, versioning it and deploying it to a fleet of cameras. It is not a camera and not a vision tool. Cognex put it in beta in June 2025 and made it generally available on 13 May 2026, and states that inspection runs entirely on the edge once a model is deployed. The cloud is where models are built and governed, which is why the useful questions for a plant are about data and networks, not image quality. This page covers what Cognex says it does, what is not public, and what to ask. It belongs to Cognex deep learning and AI, in the Cognex hub.

Illustration of a purple cloud with server modules connected to a screen on a dark background.
Training and governance sit in the cloud. Cognex states that the inspection itself keeps running on the device.

What Cognex states OneVision does

From Cognex’s general availability release (13 May 2026) and its executive’s interview with Vision Systems Design:

  • Cloud-hosted development. Labelling, training, validation and versioning of AI models happen in the cloud.
  • Fleet deployment. A validated model version can be sent to many cameras.
  • Edge runtime. After deployment the camera inspects without a cloud connection, and Cognex states images stay local.
  • Beta scale. Cognex says more than 100 customers have used it since the June 2025 beta launch.

Cognex also makes performance claims. It states that OneVision reduces scaling costs by “up to 50%”. Its release quotes Essity as saying a sealing inspection application that had taken more than a year was built in under a day, and Schneider Electric as saying yield doubled with fewer false rejects. These are Cognex’s and its customers’ statements, not independent measurements, and the reference conditions are not given.

Which cameras, and what is not public

ItemWhat Cognex statesWhat we could not find
Beta (June 2025)Select customers using In-Sight 3800 and 8900; broader product support planned for early 2026Number of beta sites beyond “more than 100” at general availability
General availability (13 May 2026)“Optimized to work with Cognex’s latest systems, including the In-Sight 3900 and In-Sight 6900”Whether 3800 and 8900 remain supported; the two releases read differently
Price and licenceNothing in the releasesSubscription or perpetual, per camera or per model
Regions and hostingNothing in the releasesWhere data is stored; regional availability
Model portabilityNothing in the releasesWhether trained models can be exported or used outside the platform
Relation to VisionPro Deep LearningNot statedWhether PC-trained projects can move across

The gaps are most of the table. In our view that is normal for a product six months past launch, but a plant should treat each blank cell as a question for Cognex, in writing, before it commits.

OneVision and the AI lifecycle

OneVision model lifecycle across cloud and plant Two zones. The top zone is the cloud, where OneVision hosts three steps: label and train, validate, and version. An arrow runs from version down to the plant zone, labelled deploy the approved model. In the plant zone a camera runs the model on the edge and sends results to the PLC. A dashed return path from the plant to the cloud carries drift, rejects and new images, labelled with the question of what leaves the plant. A gate on the deploy path says deployment happens only on plant approval. Cloud: OneVision 1. Label and train your images, your classes 2. Validate test set, known rejects 3. Version numbered, approved model Plant: edge and OT network 4. Camera runs model no cloud needed, per Cognex 5. PLC acts on result reject, stop, log Deploy only on plant approval Drift, rejects, new images: what leaves the plant?
The loop only works if new images flow back to the cloud and approved models flow down. Both paths cross your network, and both are governance questions.

Our read. The value Cognex describes is operational: many cameras, one governed model history. The risk is the same thing seen from the plant: a model pushed to a line is a change to a validated process, and the loop has two network paths that someone has to own.

IT/OT and data-governance questions

Ask these before a pilot, and get the answers in the contract or a technical document, not on a call.

  1. What is uploaded? Training images, labels, inspection results, camera diagnostics, all or none of these. Cognex states images stay local at runtime, but training needs images in the cloud.
  2. Where is it stored, and who can see it? Region, tenant separation, Cognex staff access, sub-processors, retention and deletion on request.
  3. Who owns the models and the data? Can you export a trained model, and what happens to it if you leave?
  4. What network paths does it need? Which direction connections start, which endpoints, which ports. In our view an outbound-only connection from a camera or gateway is easier to defend than inbound access. See OT network architecture for small and medium manufacturers and ISA/IEC 62443 for the zone and conduit thinking.
  5. Does a camera need the cloud at deployment only, or also at runtime? Cognex states runtime needs no cloud. Verify it by cutting the connection in a test.
  6. Who approves a model push to production? For a regulated or audited line, deployment must go through change control, with a rollback to the previous version.
  7. What does the audit trail record? Who labelled, who approved, which version ran on which camera on which day.
  8. What happens if the service is down or the contract ends? Cameras should keep running the last approved model.
  9. Is there an air-gapped route? Some plants cannot connect production cameras to a cloud at all.

On the camera side, Vision Suite release history lists features that bear on these questions: HTTPS (25.1.0), audit logging (24.3.0), single sign-on (24.4.0), a Services and Ports setting (25.3.0), MQTT enterprise connectivity (25.3.0) and a minimum password strength option (26.2.0). They are camera controls, not OneVision controls.

A sensible pilot

In our view a OneVision pilot is as much an IT and quality exercise as a vision one. A sequence we would use:

  1. Pick one line where a false reject is cheap and the data is not sensitive.
  2. Classify the images. Do they show people, recipes, branded product or anything covered by a customer agreement? That decides what may leave the plant.
  3. Test the network path with IT: outbound only, through the approved proxy or gateway, and logged.
  4. Rehearse the push and the rollback. Deploy a model, then return to the previous one, and time both.
  5. Cut the connection with the line running and confirm the camera keeps inspecting.
  6. Test the exit. Ask what you can take with you if you stop, before you have a hundred models in the platform.

One related point on licensing: per Cognex’s support documentation as we read it, Edge AI in the Vision Suite emulator is licensed through a cloud credential file, and the PC needs internet access to import it. That is a separate dependency from OneVision, but it affects offline engineering stations in a locked-down plant.

Where it gets hard

The blanks. Price, hosting region, supported cameras after launch and model portability are not public, so a business case rests on assumptions until Cognex answers.

Two systems, one change record. The cloud holds the model history and the PLC holds the recipe. Keep a single record of which model was live when a batch was run.

Validation. In a regulated plant, cloud-managed models still need your own qualification evidence. We could not confirm from public sources what OneVision provides for Part 11, so ask for it in writing. See the validation discussion on the parent page.

Vendor dependence. A model that exists only in a platform you rent is a dependency. In our view weigh that the way you would any cloud service that sits behind a production line.

For plant data paths see plant data. If you want an independent view of the data and network side before a pilot, our owner’s engineer service is where this sits. We have no commercial relationship with Cognex and do not resell its products. For the PC-based alternative, see the sibling page on VisionPro Deep Learning.

Frequently asked questions

What is Cognex OneVision?

OneVision is a Cognex cloud platform for the AI model lifecycle: developing, labelling, validating, versioning and deploying models to a fleet of cameras. It is not a camera. Cognex states the inspection itself runs on the device at the edge, with no cloud connection needed after deployment and images staying local. It reached general availability on 13 May 2026.

Which Cognex cameras work with OneVision?

Cognex's June 2025 beta release named the In-Sight 3800 and 8900. Its 13 May 2026 general availability release says OneVision is optimized to work with Cognex's latest systems, including the In-Sight 3900 and In-Sight 6900. The two releases do not make clear whether the 3800 and 8900 remain supported, so confirm with Cognex before you plan a fleet around either.

How much does Cognex OneVision cost?

Cognex has not published pricing, a licence model or regional availability for OneVision in the releases we could read. Treat cost as an open question for your Cognex representative, and ask whether it is a subscription, whether it is per camera or per model, and what happens to your models if you stop paying.

Does OneVision send my inspection images to the cloud?

Cognex states that inspection runs on the edge and images stay local at runtime. Training is different: models are developed and labelled in the cloud, so training images must reach it in some form. Ask Cognex exactly what is uploaded, where it is stored, who can see it and how it is deleted.

Is OneVision a replacement for VisionPro Deep Learning?

Not as Cognex describes it. OneVision manages the lifecycle of AI models for edge systems, while VisionPro Deep Learning is PC and GPU software. Cognex has not published how the two relate, so we do not claim either replaces the other. Check with Cognex before assuming a VisionPro Deep Learning project can move into OneVision.

References and further reading

  1. Cognex Launches OneVision (general availability) · Cognex (via AAP)The 13 May 2026 release: the model, the claims and the customer quotes.
  2. Cognex OneVision beta announcement · Cognex (via AAP)June 2025: select customers, In-Sight 3800 and 8900.
  3. Cognex's Brian Benoit answers our questions about the In-Sight 6900 vision controller · Vision Systems DesignA Cognex executive on how OneVision and the edge fit together.
  4. In-Sight Vision Suite EasyBuilder release history · Cognex documentationCamera-side security and connectivity features, listed by release.
  5. Cognex launches highest-performance embedded AI vision system powered by Qualcomm · Cognex (via AAP)The In-Sight 3900 release, which pairs the camera with OneVision.

← Deep Learning and AI

Have a project that needs technical ownership?

Send us the situation. If it is not something we should take on, we will say so.

The Plant-Side Brief

The weekly brief for the people who own a plant’s control systems.

One email every Wednesday for engineering and operations leaders responsible for controls, SCADA, MES and OT networks. Skim the top in two minutes; read the Deep Dive when you have fifteen.

  • One modernization problem, worked through to cost, risk and a schedule you can defend. The Deep Dive.
  • End-of-support dates before they reach your plant Windows, servers, controllers and HMI software. Lifecycle Watch.
  • A checklist your engineers can run that week not a framework to admire. Field Kit.
  • Every number sourced, no vendor agenda we are paid by plants, never by vendors.