By August 2026, Google updated its Chrome support documentation to reveal that the browser requires approximately 20GB of free disk space before it will even begin downloading its on-device generative AI models . Microsoft Edge follows a similar pattern. This article explains what the 20GB figure really means, how much storage the actual models consume, how to opt out as an individual or enterprise admin, and why the practice sparked such a backlash.
The 20GB figure is a free-space gate, not the size of the AI model itself . Both Chrome and Edge require roughly 20GB of available disk space as a precondition before they will initiate the download of on-device generative AI models (Gemini Nano for Chrome, a comparable foundational model for Edge) .
Google's official support documentation lists approximately 20GB of free disk space, an unmetered network connection, and sufficient device performance as eligibility requirements . The high threshold is a safety buffer to ensure the download, decompression, and temporary staging files have enough room to complete without failures or performance degradation . If a device has less than 20GB free, Chrome simply skips the download entirely .
Once downloaded and installed, the actual storage footprint is much smaller:
| Source | Post-Installation Size |
|---|---|
| Google's official confirmation (via Android Authority) | ~4 GB |
| Independent reporting (Ubergizmo, Tom's Hardware) | ~4 GB |
| Individual user observations (weights.bin file) | 4.0–4.3 GB |
| Microsoft Edge (EdgeLLMOnDeviceModel) | ~2.3–2.5 GB |
Chrome's core model file is named weights.bin and lives inside the OptGuideOnDeviceModel folder within the browser's user data directory . Edge stores its analog under %LOCALAPPDATA%\Microsoft\Edge\User Data\EdgeLLMOnDeviceModel\ .
A caveat: Chrome's background updater has been reported to sometimes fail to delete older model versions, which can cause the folder to balloon to 12GB or more on affected systems .
You can check the current status and size of Chrome's on-device AI model by opening this URL in Chrome:
chrome://on-device-internals
This internal page shows model status, version, and storage details . You can also check the file directly in your Chrome user data folder under OptGuideOnDeviceModel .
Google added an official toggle in Chrome's settings in February 2026 :
Turning this off deletes any downloaded model files and prevents future downloads . Users can also disable related AI feature flags via chrome://flags for more granular control .
Warning: Manually deleting the weights.bin file is ineffective — Chrome will simply re-download it on the next restart .
Microsoft Edge does not offer a simple user-facing toggle for the local foundational model, but users can disable it through registry settings or policy (see enterprise section below) .
Key enterprise policies:
1 (Allow feature without improving AI models) or 2 (Disable feature) .1 (Do not allow) to disable all Gemini integrations across the organization .Microsoft provides a GenAILocalFoundationalModelSettings policy that works for both Edge and Chrome on Windows :
To set via Registry Editor:
For Edge: HKEY_LOCAL_MACHINE\SOFTWARE\Policies\Microsoft\Edge
For Chrome: HKLM\SOFTWARE\Policies\Google\Chrome
Create a DWORD (32-bit) value named GenAILocalFoundationalModelSettings and set it to 1 .
To set via Command Prompt (admin):
# For Chrome
reg add "HKLM\SOFTWARE\Policies\Google\Chrome" /v "GenAILocalFoundationalModelSettings" /t REG_DWORD /d 1 /f
# For Edge
reg add "HKLM\SOFTWARE\Policies\Microsoft\Edge" /v "GenAILocalFoundationalModelSettings" /t REG_DWORD /d 1 /f
On macOS and Linux, equivalent configuration files apply. See Microsoft's and Google's enterprise documentation for details .
Several factors combined to make this one of the most contentious browser privacy stories of 2026:
Lack of consent. The download happened automatically in the background on eligible machines without a clear opt-in prompt .
Self-reinstalling behavior. Users who deleted the file reported that Chrome re-downloaded it on the next restart, making manual removal futile .
Privacy documentation quietly removed. Chrome 148 (May 5, 2026) removed the claim that on-device AI "allows you to use AI models that run directly on your device without sending data to Google servers" — raising concerns about what data might now be sent to Google's servers . Google's spokesperson later told Decrypt that the removal "does not reflect a change in how on-device AI is processed" and that data is still processed locally .
W3C standards controversy. Google shipped the Prompt API (a JavaScript interface to Gemini Nano) over formal objections from Mozilla, Apple's WebKit team, the W3C Technical Architecture Group, Microsoft, and even a Chrome engineer who had helped approve the feature .
EU legal risk. Alexander Hanff argued that the silent installation may violate the EU's ePrivacy Directive and GDPR .
Energy and bandwidth costs. The 4GB download across millions of machines was estimated to waste significant bandwidth and electricity .
Snopes later fact-checked the claims and largely confirmed the core behavior (silent download, self-reinstallation upon deletion), rating the core assertion as "mostly accurate" .
The 20GB requirement is a download gate, not a permanent storage tax — the actual Gemini Nano model takes up about 4GB. But for many users, the real issue isn't the disk space: it's the lack of transparency and consent. Whether you're a privacy-conscious individual or an IT administrator managing a fleet of machines, the opt-out methods above give you control over what lands on your SSD.