Google Documents Chrome's 20GB On-Device AI Requirement
Google's Chrome support documentation now says eligible devices need approximately 20GB of free disk space, an unmetered connection, and sufficient performance before Chrome downloads its on-device generative AI models. Google separately told Android Authority that the installed models still occupy about 4GB, making the 20GB figure a download prerequisite rather than the steady-state model footprint.
Google's Chrome support documentation now specifies that eligible devices need approximately 20GB of free disk space before Chrome downloads its on-device generative AI models. The page lists two other download conditions: an unmetered network connection and sufficient device performance.
Chrome may download the models in the background or when a user first invokes a feature that requires them, according to Google's support page. Google lists writing and rephrasing assistance, scam warnings, webpage summaries, and tab organization as examples of features that may use local models.
Eligibility and model management
Google says Chrome downloads the models only on devices with supported hardware capabilities. It does not publish a fixed hardware list on the page and warns that the specific requirements can change over time. PiunikaWeb independently retrieved the updated requirements and documented the same storage, network, and performance conditions.
The setting is controlled per user and per device. Google directs users to Settings > System, where the On-device AI toggle can be enabled or disabled. Turning the setting off deletes the local model files and prevents another download until the setting is re-enabled. Google advises against manually deleting model files because that may not prevent Chrome from downloading them again.
The installed models are about 4GB
The 20GB figure is a free-space gate for the download, not the models' steady-state footprint. Google told Android Authority that the models themselves still occupy approximately 4GB once installed. The company did not explain why the download prerequisite is roughly five times larger, leaving open whether staging, updates, rollback capacity, or another operational factor accounts for the difference.
That distinction matters for endpoint management. A laptop may need 20GB available to qualify even though the retained artifact is much smaller, and a device below the threshold may not receive local-model features at all. Teams should therefore track both the temporary eligibility headroom and the expected steady-state storage use.
Google describes local execution as processing data directly on the device, which it says can improve privacy and speed. Not every Chrome AI feature relies on these downloadable models: the support page says some AI features may continue working after local models are removed or the setting is disabled.
For enterprise and endpoint-management teams, the updated documentation makes storage capacity an explicit deployment constraint alongside connectivity and hardware performance. The remaining unknown is not the approximate installed size, which Google gave as 4GB, but the reason for the larger 20GB prerequisite and the precise hardware configurations Chrome considers sufficient.
Key Points
- 1Chrome requires approximately 20GB of available storage before eligible devices can download its on-device generative AI models.
- 2Google told Android Authority that the installed models occupy about 4GB, so the 20GB figure is a download prerequisite rather than the final footprint.
- 3Local-model rollouts make temporary storage headroom, network policy, hardware eligibility, and feature fallbacks relevant to endpoint management.
Scoring Rationale
The documentation gives practitioners and endpoint administrators a concrete storage prerequisite and an approximate installed footprint for Chrome's local AI features. It is operationally useful for device fleets and storage-constrained laptops, but it does not introduce a new model, API, or capability.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
