Opera wants the AI features on web pages to run on the reader's own machine rather than in someone else's data center. On October 27 the Norwegian browser maker began shipping WebAI, an interface through which sites and browser extensions can call a locally hosted model, first as an opt-in preview in Opera One, the company's main browser. Opera GX, the company's gaming browser, and Opera Neon, its agent-focused browser, are slated to follow.
Developers get three building blocks, which Opera names the Prompt, Decision and Semantic Embeddings APIs. Between them they cover writing text, sorting or labelling content, and measuring how closely two pieces of information are related. Inference happens inside the browser, so a developer does not have to run a server, hand out an API key or pay a cloud provider for each call. Opera's sample uses are deliberately small: an extension that tidies a crowded tab bar, one that flags clickbait, another that picks out relevant job ads.
The implementation is not locked to a single model. Opera says users can already rely on Google's Gemma open-weight models and, on Apple silicon Macs, on Apple's on-device foundation model. The company also describes WebAI as a proposed web standard backed by firms such as Google and Microsoft. That description is Opera's own: the materials reviewed for this story contained no statement from either company and no link to a standards-body draft, and TechCrunch's report likewise relied on Opera's account.
Users decide whether any of it runs. No model is downloaded until the feature is switched on under Settings, then Opera AI, where a toggle called Enable Web AI API fetches a compatible model. Per-site permissions then control which pages may reach it. For now the preview covers 64-bit Windows machines on Intel or AMD chips and Apple silicon Macs, while Intel-based Macs, Linux and Windows on Arm are left out. Opera's notes for version 139, which moves the browser to Chromium 155, list the feature among the release's additions.
The launch builds on a 2024 experiment in which Opera gave developers access to local models in a test build. The difference now is reach. The capability is moving into the browser that ordinary users install, and outside developers get a supported way to build on it rather than depending only on features Opera ships itself. Opera says further open models tuned for different devices, plus more developer tooling, are coming.
Privacy is the core of the pitch. Because the model sits on the device, page content handled by WebAI can stay off remote AI servers, and no separate AI subscription is required. Digital Trends noted that the per-site permission step carries real weight here, since third-party websites, and not only Opera, will be able to invoke the model on a visitor's computer.
The limits are practical ones. Speed and output quality depend on the user's hardware, and Opera has not published performance requirements beyond the list of supported platforms. A compact on-device model will also trail large cloud systems on many tasks. Whether WebAI becomes a genuine cross-browser standard, rather than an Opera feature, depends on other browser makers adopting it, which has not happened yet. Opera shipped a separate experiment alongside it, HTML-in-Canvas, which places working web pages inside 2D and 3D scenes.