Opera Brings Local AI to the Browser With New Web Standard

The race to put generative intelligence into every software application usually involves sending your data straight to a massive, remote server farm. Browser maker Opera is taking a radically different approach by moving the heavy lifting directly back to your laptop. The company recently rolled out an early preview of a proposed web standard called WebAI, allowing websites and extensions to execute machine learning models locally without touching the cloud.

Traditionally, integrating smart features into a webpage requires developers to ping external APIs, rack up token costs, and hand over user data to third-party servers. WebAI fundamentally changes that dynamic by baking an inference engine right into the browser shell.

Because the processing happens on-device, the system inherently solves the biggest headache in modern tech: privacy. When you ask a cloud-based chatbot to summarize a sensitive financial document or draft an email, that text lives on a corporate server. With Opera’s new framework, the data never leaves your machine. Krystian Kolondra, the EVP of Browsers and AI at the company, noted that a vast majority of everyday machine learning tasks simply do not require a massive data center when modern consumer hardware is more than capable of handling the workload. This local-first architecture also means extensions can react in real-time to the exact webpage you are currently viewing, parsing the DOM without shipping the HTML payload across the internet.

Opera is not trying to build a walled garden here. The framework is designed to play nicely with various open-weight architectures. Right now, the implementation fully supports lightweight systems like Google’s Gemma family, and on compatible Macs, it can even tap directly into Apple’s native foundation models.

To give builders actual tools to work with, the standard introduces three distinct interfaces: Prompt for generating text, Decision for classifying content, and Semantic Embeddings for understanding contextual relationships between different pieces of information. This opens the door for indie developers to code highly personalized extensions, like a tool that automatically highlights job postings on a scrolling feed, filters out clickbait headlines based on your reading history, or generates a quick reading comprehension quiz for a long-form news article without ever worrying about server maintenance or API billing limits.

Running neural networks on your own silicon is not entirely free, though the cost is paid in battery life and thermal output rather than monthly subscriptions. The initial rollout is currently available to Opera One users running 64-bit Windows machines with Intel or AMD chips, as well as Macs equipped with Apple silicon.

Users must explicitly opt into the feature via the settings menu, which triggers a one-time download of the required model weights. While the processing speed will naturally depend on the physical hardware sitting on your desk, giving the community this level of granular control represents a massive shift. Instead of renting intelligence by the millisecond, the web is finally starting to let you own it.