Chrome 148 adds lazy media loading and a browser Prompt API

Chrome 148 started rolling out on May 5 with native lazy loading for audio and video, name-only CSS container queries, and a Prompt API for a browser-provided on-device model. The Chrome release summary says the prompt implementation accepts text, image, and audio inputs and can constrain output with regular expressions or JSON Schema.

Native media deferral removes common glue code

Developers can add loading="lazy" to <video> and <audio> so resources are deferred until the element approaches the viewport. This follows the established attribute used by images and iframes. It can reduce initial transfer on pages with long media lists, while leaving loading decisions to the browser instead of an IntersectionObserver wrapper.

The important boundary is user intent. A podcast player expected to continue in the background, a hero video needed immediately, or media used for layout metadata may not be a good lazy-loading candidate. Test seeking, poster rendering, autoplay policy, and controls before applying the attribute across a component library.

Name-only container queries let @container --name match a declared container-name without also assigning a container type. This makes it easier to route styles through component boundaries when size or style containment is unnecessary.

On-device prompting needs capability-aware UX

The Prompt API opens local classification and structured-extraction experiments without an application-managed model endpoint. Availability can still depend on browser version, hardware, model download, policy, and language support. Applications need a capability check and a useful path when the model is unavailable or output fails its constraint.

Keep the first use case narrow and observable. Record whether the task completed, was rejected, or fell back, without collecting sensitive prompt data by default. Validate JSON as untrusted input even when a schema constraint is requested. Chrome 148 is a platform release, not a reason to make a core workflow depend exclusively on one browser’s model runtime.