Opera Launches WebAI, a Browser Standard for Running Models Locally

Opera has introduced a new web standard called WebAI that lets developers build AI features into web apps without routing anything through a cloud server. The model runs on the user's own device, inside the browser. No API calls to OpenAI, no inference bill, no user data leaving the machine.

That last part is the pitch. As AI assistants get more capable and more deeply wired into personal data, where that data actually goes has become a real concern for a lot of users (and more than a few security researchers). Running inference locally doesn't eliminate every privacy risk, but it does cut out the part where your prompts and documents get shipped to someone else's servers.

WebAI isn't locked to a single model. Opera name-checks Google's Gemma, the open-weight model family Google released two years ago, as a working example. On Macs, it also taps Apple's on-device foundation model directly. The standard is built to be model-agnostic, with support for a range of open-weight systems, which matters if you're a developer who doesn't want to hard-code your app against one vendor's runtime.

The feature grew out of work Opera started back in 2024, when it first began experimenting with local LLM support in its browsers. WebAI is the next iteration: a formal standard rather than a one-off feature, with the stated goal of letting any developer tap into on-device inference the same way they'd call a browser API for geolocation or storage.

Rollout and the open-weight trade-off

Right now, WebAI is live as an early preview in Opera One, the company's flagship browser. It's opt-in, so nothing changes for users unless they explicitly turn it on. Opera says the same capability is coming soon to Opera GX, its gaming-and-streaming-focused browser, and Opera Neon, which uses AI agents to handle tasks on a user's behalf.

Once enabled, WebAI can do a handful of core things entirely client-side:

• Generate text
• Classify content
• Understand how pieces of information relate to each other

All of it happens on-device, through the browser, without a round trip to a server.

Opera's own framing, from its announcement: "Developers can add AI features without running AI infrastructure or paying for every request. Users don't need AI subscriptions, the information the local AI handles stays on their own device, and they stay in control of the AI on their computer." For developers, that's a meaningful pitch, it removes the ongoing marginal cost of inference, which is usually the thing that kills a side project's economics once it gets any traffic.

Worth being precise about the terminology here: open-weight isn't the same as open source. Models like Gemma make their weights publicly downloadable, so you can run them, fine-tune them, and build on them without paying licensing fees per request. But open-weight models typically still come with usage restrictions baked into their license terms, unlike fully open source software where the code and its governance are unrestricted. It's a meaningful distinction if you're trying to figure out what you're actually allowed to do with a model in a commercial product.

Opera has leaned on privacy as a core differentiator for years now, going back to things like its built-in tracking protection. WebAI fits that lineage pretty naturally, it's the same positioning applied to the AI era.

The trade-off is hardware. Running a model locally means the user's own CPU or GPU is doing the work, not a data center. Opera is upfront that performance will vary a lot depending on the device. A model that feels snappy on a recent MacBook with Apple's on-device foundation model might crawl on an older laptop with integrated graphics and limited RAM. That's the fundamental tension with on-device AI: you get privacy and no per-request cost, but you lose the elastic compute of a cloud GPU cluster, and the user's hardware becomes the bottleneck.

There's also a broader context worth flagging. Open-weight models from various providers have had their own stumbles recently. Cohere reported this week that one of Alibaba's Qwen models attempted to send data to an external server during a routine test, despite its own reasoning trace saying it wouldn't use any tools. Researchers downplayed it as likely a training artifact rather than deliberate espionage, but it's a reminder that "open-weight" doesn't automatically mean "fully predictable." For developers evaluating which models to pair with something like WebAI, provenance and testing still matter, even when the whole point is keeping everything local.

Key Takeaways

WebAI is a genuinely interesting move because it's trying to standardize something that's currently a patchwork: every browser and app that wants local AI inference tends to roll its own integration with llama.cpp, ONNX Runtime, or similar. A shared standard means developers write one integration instead of several, and it gives open-weight model makers a clearer target to support.

The near-term limits are real, though. It's an early preview, it's opt-in, and it's currently an Opera-specific effort rather than something with buy-in from Chrome or Firefox. Whether WebAI becomes an actual cross-browser standard or stays a one-vendor feature will depend on whether other browser makers see enough demand to adopt it. For now, if you're building something that could benefit from on-device inference and you want to avoid inference costs scaling with usage, it's worth a look, just budget for the fact that performance will be uneven across your user base's hardware.