Chrome's Gemini Nano Download: What Google's Docs Say, and the Numbers They Leave Out
Software / explainer
Chrome's Gemini Nano Download: What Google's Docs Say, and the Numbers They Leave Out
Google's page on built-in model management gives triggers, a 30-day rule and full-download updates, but no file size, no disk threshold and no consumer off switch.
Chrome fetches and updates its on-device Gemini Nano model without asking, and Google's developer documentation does not say how large the file is, how often it ships or how an ordinary user turns it off. A developer who runs Perishable Press says that gap is costing him bandwidth.
Jeff Starr, who describes himself as a web developer and book author, updated his post on November 19. He wrote that his network meter showed Chrome using "literally hundreds of GBs per month" until he disabled everything under the settings labels "AI innovations" and "On-device". He names no Chrome version and no monitoring tool, and gives no per-session figures. Treat the number as one developer's report, not a measurement.
What triggers the Gemini Nano download in Chrome
Google's page, Understand built-in model management in Chrome, was published October 21, 2025. It says the first call to a create() function of any built-in AI API that depends on Gemini Nano starts the download, with Summarizer.create() as its example.
A second trigger needs no page script at all. The page says availability() can start the download shortly after a fresh profile launches, if the Gemini Nano-powered scam detection feature is active.
Before downloading, Chrome estimates GPU performance with a representative shader. It then fetches a 4B-parameter variant, a 2B-parameter variant, or falls back to CPU inference. A device that fails the hardware requirements gets no model.

Updates are full downloads, checked at startup
Chrome checks for Gemini Nano updates when the browser starts, and checks supplementary files such as LoRA weights daily. The page says each update is a full download, not a delta. Updates are hot-swapped with no downtime.
LoRA weights, small add-on adapters, come down with the base model for the APIs that use them, the Proofreader API among them. Other APIs do not get them ahead of time.
| Event | What the page says | Number given |
|---|---|---|
| First download | First create() call, or availability() with scam detection on | None |
| Interrupted download | Resumes; browser must reopen in time | 30 days |
| Update check | At browser start; LoRA files daily | Daily |
| Update size | Full download, not a delta | None |
| Disk pressure | Model deleted below "a certain threshold" | None |
| Ineligible for a period | Model purged | 30 days |
What the page does not give
The page has no file size in GB or MB, and no free-disk threshold. It does not list which APIs use which model. It has no guidance on metered connections. The only way to disable the feature that it names is an enterprise policy.
That sits oddly beside Starr's account of consumer settings called "AI innovations" and "On-device". The documentation does not mention those labels, so it neither confirms nor contradicts him. The practical gap is for a reader on a capped plan, who cannot learn from Google's page what one download costs.
What Starr did to stop it
His fix has four steps: open Chrome's settings, find the "AI innovations" section, disable everything on that page, then disable the "On-device" settings. He says the meter went from spinning to "virtually flatline", and that he is now switching to other browsers and rediscovering Vivaldi.
He also wrote that "Chrome should have left the AI resource hogs disabled by default." The documentation describes the opposite default for eligible devices: the download starts on the first qualifying API call or, with scam detection on, on a fresh profile. The same kind of quiet, background work shows up elsewhere in desktop software, as in the Windows 11 coding-agent telemetry report, where the question is also what a user was told and when.
What would settle the bandwidth question
A purge, per the page, can happen mid-session and is not followed by an automatic re-download. A purge-and-fetch loop is therefore not the behaviour the documentation describes. Whether Starr's meter was counting model updates, something else in Chrome, or a different process is not something his post settles.
The documentation does give one way to look: chrome://on-device-internals shows which model version is installed. A version number that changes between daily checks would point at update churn, and one that holds still would point elsewhere. The page adds that JavaScript cannot query the model version, so the check has to be done by hand. Google's separate Android checkout work is covered elsewhere on this site.
Sources
More in Software
- 01OpenAI's WebRTC Voice Stack Draws a 30 ms Objection That Its Own Engineers Never AddressedLuke Curley says WebRTC drops audio older than 30 ms. OpenAI's engineers wrote about port ranges and routing, which is a different problem.
- 02Superpowers Passes 322,000 GitHub Stars as a Set of 15 Prompted Skills for 16 Coding AgentsJesse Vincent's MIT-licensed framework tells agents to brainstorm, plan and test first. Its clone count is large, and its evidence for better results is one author's view.
- 03GitHub Disables 144 Picofly Repositories After a Nintendo Notice Names OneThe November 6 notice is the first on GitHub aimed at Switch modchip firmware, and it asked for forks to be taken down without listing a single one.
- 04Zerodha's Blobasaur Stores Kite Layouts in Sharded SQLite, and Gives Up Durability to Do ItThe Rust, AGPLv3 blob store speaks the Redis protocol, shrank sampled production writes about 11 times, and has no live rebalancing or automatic failover yet.