{"slug": "which-ai-tasks-can-your-website-run-in-edge-and-chrome", "title": "Which AI Tasks Can Your Website Run in Edge and Chrome?", "summary": "Microsoft opened an origin trial on November 19 letting registered websites use Edge's built-in Aion Instruct model through the browser's Prompt API, while Google lists Chrome's Gemini Nano-backed Prompt API as available to websites from Chrome 148 and to extensions from Chrome 138. Both APIs share the LanguageModel JavaScript interface, but local inference still depends on a supported browser, the right hardware and an installed model, so developers must check readiness, prompt users before a model download, and offer an explicit opt-in before sending text to a hosted service.", "body_md": "A product manager pastes an unpublished launch brief into a website and asks for three bullets to send to a colleague. The brief names a partner under embargo. If the website can summarize it on the laptop, the manager can review the bullets without sending the text to a model provider. That is possible today on some Edge and Chrome installations. On another visitor’s phone, the very same button may have no local model to call.\n\nOn November 19, Microsoft opened an origin trial for sites to use Edge’s built-in Aion Instruct model. An origin trial is temporary access to a browser feature for registered sites. Chrome already offers ordinary websites a Prompt API backed by Gemini Nano. A shared API name gives developers a place to start. Visitors still need a supported browser, the right hardware and an installed model. For the person holding an unpublished brief, the decisive question is what the button does when any of those conditions fails.\n\n1. 01Edge websites join a trialAion Instruct is available as Edge’s built-in model. Register the production site for the Prompt API origin trial before relying on access from visitors’ browsers.\n2. 02Chrome websites can use the core APIGoogle lists Chrome 148 for websites and 138 for extensions. Device requirements still decide whether a particular user can run Gemini Nano.\n3. 03First use can need a downloadCheck readiness for the actual task, prompt the user before starting a download, and let them keep editing while the browser obtains its model.\n4. 04Cloud processing needs its own choiceKeep the document usable when a local request fails. Explain the destination and let the visitor decline before sending any selected text for a hosted summary.\n\n## 01 — Browser supportWhich visitors can use the *Prompt API?*\n\nA website can call a model supplied by the browser through a JavaScript interface called `LanguageModel`. Think of that interface as a request to use a capability already offered by the visitor’s browser. The website supplies instructions and the selected text. The browser manages the model and runs it on the visitor’s device. That arrangement avoids asking every website to choose a local runtime, serve its own copy of the model and keep it updated.\n\nFor Edge, Microsoft made Aion Instruct generally available as its built-in model while keeping a site’s access to the Prompt API experimental. Register the site for the time-limited Prompt API origin trial and deliver its token on the registered origin. Microsoft supports an HTML meta tag or an HTTP response header. A successful test on a developer’s machine with a browser flag enabled leaves this production step untested. Visit the deployed page in an ordinary profile before inviting customers to use it. [Microsoft’s Prompt API guide](https://learn.microsoft.com/en-us/microsoft-edge/web-platform/prompt-api)\n\nGoogle lists the core Prompt API as available to Chrome websites from version 148 and to extensions from version 138. An ordinary site targeting Chrome 148 or later can use the core API on eligible devices without building an extension or registering for the separate experiment in web sampling parameters. Those parameters adjust how varied a model’s answers can be. A team drafting three bullets from a brief can start with the core API and evaluate its answers before it needs that degree of control. [Google’s Chrome Prompt API guide](https://developer.chrome.com/docs/ai/prompt-api)\n\n## 02 — Feature choiceStart with work the visitor can check\n\nGive the browser a short excerpt from the brief and ask for the audience, the release date as written, and one unresolved dependency. The expected answer can sit next to the source. A product manager can quickly spot an invented partner approval or a date stripped of its “tentative” qualifier. A bounded task like this makes a good first local feature. It is clear where the evidence comes from and who reviews the result.\n\nOther sensible starting points are sorting a paragraph into one of a few editorial categories, shortening a selected passage, or extracting named fields from a note. Microsoft describes Aion Instruct as a 2.3-billion-parameter model for text tasks, including structured output. Google also documents constrained responses, which let the site ask for a particular output shape. A valid field called “launch date” can still contain a wrong date. Preserve the original beside the extracted value and let the author correct it. The [guide to structured output reliability](https://www.digitalapplied.com/blog/llm-structured-output-json-reliability-production) explains this distinction for applications that save generated fields.\n\nChrome’s documentation also describes image and audio inputs with text output. Audio prompting requires a suitable GPU. A Chrome capability provides little evidence about how the same request will behave in Edge, where Microsoft documents Aion for text. Ship the text brief first if that is the common use case. A feature that captions a photograph attached to the brief needs its own capability check and quality evaluation. Asking a model to verify the partner’s current public announcement would need an additional source of information. The local text model only has the material the application supplies and its learned knowledge.\n\n## 03 — Hardware limitsWhy two current browsers can disagree\n\nThe device matters even after the site and browser qualify. Microsoft’s guide lists at least 20 GB free on the drive holding the Edge profile, a GPU with at least 3 GB of video memory and an unmetered or unlimited connection for the model download. It lists Windows 10 or 11, macOS 13.3 or later, and qualifying 64-bit Linux systems with Vulkan GPU drivers from Edge 159. Edge settings and enterprise policy can also prevent the model from downloading. The free-disk requirement describes room the browser wants on the drive. It is a poor substitute for the download size shown to a visitor.\n\nGoogle’s guide lists at least 22 GB free on the Chrome profile volume. For the text feature, it lists either a GPU with more than 4 GB of video memory or a CPU configuration with at least four cores and 16 GB of system memory. Supported systems include eligible Windows, macOS and Linux desktops and Chromebook Plus devices. Chrome on Android and iOS are outside the documented support for these foundation-model APIs. A test on a developer’s recent Mac therefore tells a team very little about a visitor opening the launch brief from a phone. [Chrome’s device and network requirements](https://developer.chrome.com/docs/ai/prompt-api#review-the-hardware-requirements)\n\nRun a capability check in the actual page. The absence of `LanguageModel` is enough to suppress the local-model action. If the interface exists, ask `LanguageModel.availability()` whether the specific operation is available. In Chrome, pass the language and input type required by the feature when checking availability and creating the session. A successful English text check is a narrow claim about English text. It gives the customer no reason to expect an audio recording or a document in another language will also work.\n\n## 04 — First useExplain what the download is for\n\nThe first use on an eligible device can require a model download before the first bullet appears. Edge downloads its model when a site or browser feature needs it. Chrome likewise supplies Gemini Nano separately from the browser and lists an unmetered or unlimited connection for the initial download. The download brings model files to the device. Running a local prompt then gives the installed model the selected text on that device. Those two directions of travel are worth distinguishing in the button copy.\n\nIf the model is available, show “Summarize on this device.” If it is downloadable, offer “Download model and summarize here” and explain that the original text stays in the current document during setup. When download is in progress, show its progress and leave the text editable. Chrome documents user activation for creating a session when a download is needed. Put the action behind the user’s click. Opening a shared brief on a train should leave the reader able to choose whether to spend time and bandwidth on a model.\n\nHandle a change of mind during the wait. Someone may remove the partner name before running the summary, or decide the slide deck already contains suitable bullets. Let them keep editing or abandon the request. If the model becomes ready after they change the selection, use the latest explicitly selected text or ask them to run the action again. Otherwise an apparently successful download can produce a summary of a draft the user has already revised.\n\n##### Run locally\n\nShow the selected passage and let the visitor generate the summary on this device.\n\n##### Offer the download\n\nExplain what the browser will download. Show progress while keeping the document accessible.\n\n##### Keep the editor working\n\nOffer a clearly labeled hosted option if the workspace permits it, alongside continuing without AI.\n\n## 05 — Data handlingWhat does “on this device” cover?\n\nMicrosoft and Google describe the local model invocation as processing the supplied input on the device. That is a specific promise about the model path. A website may separately save the document to its own server, capture editor content in diagnostics, or send the resulting bullets to a shared workspace. A site that sends keystrokes to a collaboration service cannot promise the whole draft never leaves the laptop simply because it runs the summary locally. Say exactly which flow stays local, and describe the document’s own save behavior where users expect to find it. [Microsoft’s explanation of local processing](https://learn.microsoft.com/en-us/microsoft-edge/web-platform/prompt-api#alternatives-to-and-benefits-of-the-prompt-api)\n\nFor an unpublished brief, the useful control is selection. Let the author highlight a passage and show what will be summarized. A local prompt need not include the partner’s contact details, the reviewer comments or the rest of the open document. Keep generated bullets in a draft area until the author deliberately inserts them. That also avoids sharing a newly generated sentence with colleagues just because the editor normally syncs every change.\n\nOnce the model files are present, both vendors describe local prompting as able to run without a network connection. Try that in the actual editor. A previously opened, usable page can continue to summarize locally if its required assets are available. A page that needs to authenticate, retrieve the source brief or load its interface from the network may fail before it ever reaches the prompt. Describe offline support only for the operations the complete product can deliver. The [on-device AI deployment guide](https://www.digitalapplied.com/blog/on-device-local-ai-agents-2026-privacy-cost-stack-forecast) covers the wider trade-offs when a team controls the software installation.\n\nA website can also embed someone else’s editor in a frame. In Chrome, Google documents the Prompt API as available by default to top-level windows and frames from the same origin. A cross-origin frame needs explicit permission from its host to use the capability. For a document product sold both as a standalone app and as a customer portal widget, check both entry points. A customer who authorizes local processing in the standalone editor may still encounter an unavailable capability in the embedded version. The host’s permission to use a browser feature and the author’s choice to send a passage to a provider are separate decisions. [Chrome’s Prompt API permissions policy](https://developer.chrome.com/docs/ai/prompt-api#permission_policy_iframes_and_web_workers)\n\n## 06 — Visitor choiceOffer a cloud summary the author can decline\n\nImagine the product manager returns on a work laptop whose organization has disabled local models. They have already pasted the confidential paragraph and clicked a button labeled “Summarize on this device.” That click authorized the local operation as described. Switching the request to a hosted service inside an error handler changes the destination for text the author may have deliberately kept off that service. The interface needs another decision from the person holding the brief.\n\nGive a specific reason in plain language, with a recovery option that fits it. A download in progress calls for progress and a way to keep editing. A device that cannot run the model can offer a hosted summary with the provider or service named and the selected text visible. Give that action a direct label such as “Send selected text for a cloud summary.” Beside it, offer “Keep editing without AI.” Preserve the text and allow manual copying even when the person declines. For a workspace that prohibits hosted processing altogether, the useful fallback is the ordinary editor.\n\nApply the same rule when the local model returns an unsatisfactory answer. Keep “try again locally” and “use the cloud” distinguishable. The first repeats the local work. The second is a new disclosure. A support team can collect the chosen route, whether the operation completed and the duration without collecting the passage. That gives product owners evidence about unmet demand while respecting the reason many customers chose the local action in the first place.\n\nThis device can’t run the local summary. You can send the selected paragraph to our hosted summarization service, or keep editing here. In a live product, name the service and show the relevant handling terms beside that choice.\n\n## 07 — Acceptance testRun one brief through the whole journey\n\nWrite a sample launch brief with a provisional date, an unnamed partner and a dependency whose owner has yet to agree. Decide in advance which facts a summary must preserve. Read the output in eligible Chrome and in Edge on the trial origin. A fluent bullet saying the partner approved the date should fail in both. Have an editor mark whether each bullet can be traced to the sample and whether a reader could mistake an open question for a settled commitment. A small model suitable for rewriting may still need narrower instructions for this extraction. [Choosing tasks for small local models](https://www.digitalapplied.com/blog/small-language-models-on-device-agents-2026-guide)\n\nRepeat the task on a fresh eligible profile, a device without the necessary hardware, and an installed model in an already open page taken offline. In a separate run, decline the cloud option, continue typing, and check the application’s network requests to confirm that the selected text stayed out of any model-service request. Use a synthetic brief for these checks. Exercise an unavailable or expired Edge trial token too, since a site can lose trial access even on a perfectly capable machine. This exposes several customer experiences that disappear if every demonstration runs on the same warmed-up laptop.\n\nFor launch planning, keep distinct counts for people offered local inference, those who chose a download, completed local summaries, explicit cloud requests and people who continued manually. Record corrections to the sample wording in the quality review. A large number of local completions tells you the feature runs. Repeated changes to the same field, especially a launch date or partner agreement, tell you whether its output is dependable enough to keep offering.\n\nThe last check is editorial. Give one colleague the original brief and another the three generated bullets. Ask each when the launch can go ahead and whether the partner has signed off. A useful summary lets the two readers give compatible answers even though one has read considerably less. If only the full brief makes the dependency clear, leave the author in control of inserting the text and change the prompt to preserve unresolved questions. This is the point of starting with a single visible task: the team can identify exactly what makes a shorter answer trustworthy and fix that before promising an assistant for the entire website. Our [web development service](https://www.digitalapplied.com/services/web-development) covers that integration across the editor, browser and hosted processing path.\n\n### Ship the local summary as a choice the site can keep\n\nStart with a short passage and an answer the author can verify. Promise on-device processing only when the requested operation can run there. A reader whose browser is ineligible should leave with the original brief intact and full control over whether a hosted service ever sees it.", "url": "https://wpnews.pro/news/which-ai-tasks-can-your-website-run-in-edge-and-chrome", "canonical_source": "https://www.digitalapplied.com/blog/on-device-ai-edge-chrome-prompt-api", "published_at": "2026-11-21 00:00:00+00:00", "updated_at": "2026-11-21 22:14:06.222559+00:00", "lang": "en", "topics": ["ai-products", "ai-tools", "developer-tools", "large-language-models", "artificial-intelligence"], "entities": ["Microsoft", "Edge", "Aion Instruct", "Google", "Chrome", "Gemini Nano", "Prompt API", "LanguageModel"], "also_reported_by": [], "takeaway": "Sites can call the browser's built-in model via the shared LanguageModel interface, but must register for Edge's time-limited Prompt API origin trial and target Chrome 148 (or 138 for extensions) while checking model readiness and download consent per visitor.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/which-ai-tasks-can-your-website-run-in-edge-and-chrome", "markdown": "https://wpnews.pro/news/which-ai-tasks-can-your-website-run-in-edge-and-chrome.md", "text": "https://wpnews.pro/news/which-ai-tasks-can-your-website-run-in-edge-and-chrome.txt", "jsonld": "https://wpnews.pro/news/which-ai-tasks-can-your-website-run-in-edge-and-chrome.jsonld"}}