Opera started shipping local AI models in its browser behind WebAI, a proposed web standard that lets websites and extensions run Gemma on your own machine
AspireAI community · Opera
Opera announced today that it has begun shipping local AI models directly inside its browsers, with the first capabilities available as an opt-in early preview in Opera One and Opera GX and Opera Neon to follow. The surface is WebAI, a proposed web standard backed by companies including Google and Microsoft, which lets websites and browser extensions reach models hosted through the browser itself.
At launch users can reach Google's Gemma and, on Macs, Apple's own on-device foundation model. Three capabilities are exposed: text generation, content classification, and identifying relationships between pieces of information. The implementation is model-agnostic and nothing is downloaded until a user turns it on, with per-site permissions deciding which websites may reach the local models. Opera ships example extensions alongside it, including one that turns an article into moving art by making similar sentences visual neighbours. WebAI is limited to 64-bit Windows on Intel or AMD and Macs with Apple silicon - not Intel Macs, Linux, or Windows on Arm.
The honest limit: this is an early preview, and because the models run on the user's own hardware, Opera states plainly that performance depends on that hardware. Opera is not building the models here; it is building the plumbing that makes someone else's open weights reachable from a web page.
https://press.opera.com/2026/10/27/opera-doubles-down-on-local-ai-shipping-models-directly-in-its-browsers/
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.