On this page
On October 27, 2026, Opera announced WebAI as an early preview in Opera One, Opera GX, and Opera Neon. Websites and extensions can call local models that run inside the browser through Prompt, Decision, and Semantic Embeddings APIs, without standing up a separate AI stack for every experiment. Opera calls this a first step toward making local AI a central browser capability, after an earlier 2024 experimental developer build that already exposed local models. For teams that already watch local-model workflows such as Cline Desktop local models via Ollama and LM Studio, the idea is familiar: keep more inference on the user’s machine, and leave cloud AI optional rather than default.
What local inference can save
Opera’s post describes WebAI as a way for sites and extensions to use AI that runs on the user’s computer. People want AI that works with the page they are already on, not just a separate chat window. Local browser APIs are meant to support that pattern.
The practical savings Opera lists are concrete:
- No separate AI app to install just to try a page-side feature
- No local server to configure (no “start Ollama / LM Studio first” step for this path)
- No AI API key to obtain before a first demo
- No per-request cloud bill for that local path, and no AI infrastructure the developer has to operate
- When page content is processed by that local AI path, Opera says it does not need to be sent to a cloud AI provider
WebAI is also opt-in. Nothing is downloaded until the user turns it on in Settings, and it can be turned off at any time. That matters for privacy messaging and for shared machines: leaving the toggle off keeps the preview model off the disk until someone chooses it.
That local story has a clear boundary. Local inference is not the same as “the whole browser session never goes online.” Model downloads, extension installs, and ordinary page requests can still use the network. A page that calls WebAI for on-device classification can still fetch remote images, analytics, or CDN assets like any other site. Performance depends on the device’s own compute; Opera notes that local AI speed is still improving and that response time varies by machine. For another Mac-side local / hybrid comparison, see how Perplexity Computer runs locally on Mac. Related on-device multimodal reading: MiniCPM-V 4.7 open multimodal on phones.
Prompt, Decision, and Semantic Embeddings
Opera names three developer surfaces in the news post:
- Prompt — generate text (the building block behind quizzes, rewrites, and short page-side answers)
- Decision — classify information (useful when a tool must pick a label, priority, or verdict rather than free-form prose)
- Semantic Embeddings — find relationships between pieces of content (similarity, clustering, “which paragraphs match this project”)
Together they are aimed at tools that do more than answer questions in a chat box: highlighting what matters on a long article, checking comprehension, ranking saved items, or mapping related sentences. Independent developers and small teams get a named nod. WebAI is presented as a way to ship personal extensions without running an external AI service for every token.
Proposed standard, not a finished universal web API
Opera describes WebAI as a proposed web standard, with backing that includes companies such as Google and Microsoft. Its implementation lets developers use a range of models in the browser, including open-weight models on Opera’s inference engine and compatible models already on the device, so the feature is not cast as a single-vendor lock-in.
Early preview here means a browser-side local AI surface available today on supported Opera builds, not a finished, industry-wide web standard that every browser ships by default. The “proposed / early preview” wording holds until a later, broader GA story exists.
How to turn it on (Opera’s three steps)
Opera’s try-it path for the preview is:
- Download or update Opera One (the same early preview is also described for Opera GX and Opera Neon)
- Open Settings → Opera AI and turn on Enable Web AI API. This downloads a compatible local model and may take a few minutes
- Install one of the example extensions and open it from the extensions menu
Until that toggle is on, WebAI stays off and does not pull models in the background. A first launch that feels “empty” usually means the toggle is still off or the model download has not finished. After the model is local, the same toggle can turn the feature off again for a one-time demo. Opera’s opt-in framing leaves that control with the user.
Devices that can run it, and those that cannot
Opera’s early preview availability is narrow enough that device choice matters before a demo or setup notes:
| Supported | Windows PC (64-bit, Intel / AMD); Apple silicon Macs |
| Not available | Intel-based Macs; Linux; Windows on Arm |
If a workflow depends on Intel Macs, Linux desktops, or Windows on Arm, this preview path is not the one to promise yet. Apple silicon support does not extend to Intel Macs; Windows support is the classic 64-bit Intel/AMD line, not Windows on Arm. Mixed fleets need the OS / CPU matrix checked before a shared trial.
Demo extensions: samples, not built-in features
To show what developers could build, Opera’s team shipped six experimental extensions. The post is explicit: they are examples to spark ideas, not Opera features. They are experimental samples, not preinstalled Opera capabilities, and removing or changing them later would not mean a core Opera feature was “deprecated.”
The six names are Project Lens, Got It?, Questionable Intelligence, Semantic Art, RecordIn Alchemy, and C·I·A·O. Two of them make the “more than a chat window” point without listing every demo in depth:
- Got It? writes a short quiz about the article just read, for an on-device understanding check
- Project Lens takes the current project and colors paragraphs by how much they matter to that project
The remaining demos lean playful or visual: charts, moving art, satire-style social posts, ASCII improvisation. They still sit on the same theme as the Prompt / Decision / Semantic Embeddings building blocks: generate text, classify information, and find relationships between pieces of content on the page already on screen. They are inspiration packs to install, not a checklist of built-in Opera AI modes.
Developer docs entry (playground overview)
Opera points developers to next.opera.com/web-ai for the supported-API list and documentation. That entry loads as a Web AI playground overview: models run on the device, and cards cover APIs such as Prompt, Summarizer, Writer, Rewriter, Translator, Proofreader, Semantic Embedder, and Decisions, plus sample applications and the demo extensions.
Product coverage here stays with the three APIs Opera’s news post names in plain language: Prompt, Decision, and Semantic Embeddings. The playground page shows which extra API cards appear in a given build. Call-by-call integration steps belong in the live docs for the Opera version and channel in use (One / GX / Neon).
Speed, Chinese / Japanese quality, and what is untested on this site
For speed and Chinese / Japanese quality, on this site there are no tests yet; feedback can go through the site-owner contact at the bottom of the site. Opera itself says response time depends on the machine, and that local AI performance is still being made lighter and faster. Until more independent reports land, the early preview is a capability to enable and explore, not a measured latency or localization benchmark.