Opera’s local-AI preview is opt-in, with hardware and site-permission limits

Opera announced a local‑AI early preview on October 27, letting websites and extensions use models hosted through the browser on a person’s own computer. The feature, called WebAI, is opt‑in. It is a preview of a capability that developers can build on, rather than a finished collection of AI tools enabled for every browser user.

The company’s two launch pages describe the browser rollout differently. Its press release says the first capabilities are available in Opera One, with Opera GX and Opera Neon to follow. Its user blog describes the preview across all three. Both document Opera One, which is also the browser named in the setup instructions. The sources do not establish identical activation timing for every GX or Neon installation.

The setup article specifies 64‑bit Windows PCs with Intel or AMD processors, and Macs with Apple silicon. It explicitly excludes Intel‑based Macs, Linux and Windows on Arm. A computer’s ability to run an Opera browser therefore does not by itself establish access to this local‑AI preview.

Opera says users enable WebAI in Settings under Opera AI. A compatible model can then be downloaded, which may take a few minutes. The company says nothing is downloaded until WebAI is turned on, and users can turn it off again. Responsiveness depends on the computer’s hardware; the announcement does not provide a measured speed, battery‑use or memory comparison for the supported machines.

Opera’s [release](https://press.opera.com/2026/10/27/opera‑doubles‑down‑on‑local‑ai‑shipping‑models‑directly‑in‑its‑browsers/) says site‑specific permissions let people choose which websites can access local AI. Its implementation supports different compatible models, including Google’s Gemma and Apple’s on‑device foundation model on Macs. The user blog describes text generation, classification and finding relationships between content as the kinds of work that sites and extensions can build. Running a model locally means that inference does not need to send the processed content to a cloud AI provider. It does not, on its own, establish that every website or extension sends no other data over the network. Model location, site access and an extension’s own behavior remain separate questions. OpenTools has not installed the preview or audited its network traffic or permission enforcement.

The user blog labels its six experimental extensions as examples, not Opera features. They include Project Lens, which highlights article paragraphs by their relevance to a project, and Got It?, which makes a quiz about an article. Those descriptions show possible uses of the interface; they are not evidence that every Opera installation includes those tools or that their outputs have been independently validated.

WebAI also builds on Opera’s earlier local‑model experiment from 2024. The October 27 change is the current opt‑in preview and its browser‑facing capabilities, rather than the invention of browser‑local AI or a universal web standard already supported everywhere.

Conceptual illustration of an on‑device model and choosing access for individual sites, created for OpenTools. It is not an Opera interface or technical architecture diagram, and does not demonstrate that a whole browser operates offline or sends no network data.