On October 27, Opera announced WebAI, a proposed web API for using local models such as Google’s Gemma from the browser. The content processed locally does not need to be sent to a cloud inference service. Opera One offers an opt-in early preview, but this is not yet a W3C standard.

What the proposal offers

Opera supports the proposed Built-in AI APIs as well as Semantic Embedder and Decisions, with local inference built into the browser. A site does not need to run its own inference server for each request, and users do not need a separate application.

Conceptual layers of an application, browser and local inference engineConceptual layers of an application, browser and local inference engine
Conceptual sketch: the API exposes local inference; implementation and hardware may vary by browser and device.

What you can measure today

WebGPU can help inspect graphics capabilities for your own experiments, but it is a different layer: the following example does not test WebAI support.

ts
const adapter = await navigator.gpu?.requestAdapter()

if (!adapter) {
  console.warn('WebGPU unavailable; this does not test WebAI')
} else {
  console.info('Max buffer bytes:', adapter.limits.maxBufferSize)
}

Limits worth measuring

  • Memory: model weights can occupy gigabytes, depending on the model and quantization.

  • Cold start: the first model download can add latency.

  • Battery: sustained local inference consumes device resources.

  • Parity: local and remote models may produce different answers to the same prompt.

Local inference changes where the computation happens; it does not eliminate the need to measure it.

Where local inference can help

For sensitive documents, local processing can reduce the need to send content to a remote provider. It does not replace a security or compliance review, and availability still depends on device support and user consent. Source: Opera’s WebAI announcement(Opens in a new tab).