Chrome Prompt API

The Chrome Prompt API is a native AI interface that calls Gemini Nano built into the browser. It is designed to perform on-device inference without server round trips, reducing latency and privacy exposure.

The key features are as follows.

• You can get a single response with prompt(), or a ReadableStream-based streaming response with promptStreaming().
• It manages session-based context, and controls session lifetime and interruption via clone(), destroy(), and AbortSignal.
• By putting a JSON Schema into responseConstraint, you can constrain the output format to a boolean or a specific JSON structure.
• Through initialPrompts, append(), and prefix: true, you can finely adjust the system prompt and the starting format of the response.
• It supports not only text but also image and audio input.
• You can specify input/output formats and languages with expectedInputs / expectedOutputs, and it currently supports en, ja, and es.

Environment requirements are also specified. It works on Windows 10+, macOS 13+, Linux, and ChromeOS, and requires 22GB or more of storage. It requires GPU VRAM over 4GB or CPU RAM 16GB or more and 4 or more cores.

Cross-origin iframes can delegate access via the allow="language-model" attribute, but web workers are not yet supported. It has been available as an origin trial since Chrome 138.