- Chrome is automatically downloading a Gemini Nano AI model of about 4GB called weights.bin into the OptGuideOnDeviceModel folder without notifying the user.
- The model runs locally for functions such as scam detection, writing assistance, and developer APIs, but it is re-downloaded even if it is manually deleted.
- Privacy and European regulatory experts question the lack of transparency, the potential clash with ePrivacy/GDPR regulations, and the environmental impact of a massive download.
- There are ways to limit or block downloads from Chrome settings, advanced flags, and policies in Windows 11, although they are not always visible or intuitive.
In recent months, many users have discovered a mysterious, multi-gigabyte file on their computers associated with Google Chrome without having installed anything unusual or changed any settings. This file, called weights.bin and saved in an internal folder of the browser, it has set off all the alarms because it occupies around 4 GB and appears without any warning or prior consent screen.
After initial suspicions that it might be malware or some type of spyware, Google has acknowledged that it is Gemini Nano integrated into ChromeIt's an artificial intelligence model designed to run directly on the device. The technical explanation hasn't exactly calmed the waters: the download happens in the background, it repeats even if the user manually deletes the file, and there's hardly any clear information on how to control or prevent it.
What exactly is Chrome downloading and where is it hiding?
What many have found on their computers is a directory called OptGuideOnDeviceModel inside the Chrome user data folder. The file appears inside. weights.binwhich, in tests conducted by researchers and users, is around 4-4,3 GB. On macOS and Linux, it is located in the browser profile, while on Windows 11 it is usually found in the path %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel.
This file corresponds to the artificial intelligence model Gemini NanoThe lightweight version of Google's Gemini models, adapted to run locally ("on-device") instead of relying exclusively on the company's servers. The goal is for certain Chrome features, such as scam detection, writing assistance, or some APIs for developers, can function directly on the user's computer without sending all the data to the cloud.
According to Google, the model is part of the browser's strategy to incorporate generative AI and security capabilities directly into the device. However, the company had barely mentioned this massive download before the privacy researcher Alexander Hanff (“That Privacy Guy”) and other specialists began to document the browser's behavior and share screenshots of the folder on social networks and specialized media.
Chrome also offers an internal page, chrome://on-device-internalswhere you can check the status of local models: installed version, exact path, disk size, and whether they are active. That section confirms that Gemini Nano It is a model downloaded and managed by the browser itself, and which is automatically updated or reinstalled when it deems necessary.
On which devices can Gemini Nano be downloaded and what functions does it use?
The download of this 4GB model is not happening indiscriminately on all computers with Chrome. Google has defined minimum hardware and system requirements so that the browser can use Gemini Nano locally. Among the conditions that have been revealed are systems such as Windows 10 or 11, macOS 13 or later, some Linux distributions, and ChromeOS on Chromebook Plus devices.
In addition to the operating system, a high level of resources is required: we are talking about computers with more than 4 GB of VRAM or, alternatively, at least 16 GB of RAM and quad-core processorsApproximately 22 GB of free space on disk so that the installation and future updates of the model can be completed smoothly, and a unmetered data connection or without strict limits, given the size of the initial download.
According to Google, the Gemini Nano in Chrome is used for tasks such as... real-time scam detection within the browseridentifying malicious pages or rapidly changing fraud attempts, as well as feeding APIs that developers can take advantage of to build functions that run entirely on the user's device. It can also intervene in writing assistance tools or contextual assistance integrated into the browser.
Paradoxically, the popular “AI mode” that some users see in the Chrome interface, with a button or capsule in the address bar, Gemini Nano is not always usedInstead, in many cases it continues to send queries to Google's servers. This creates a confusing situation: there is AI in the cloud and local AI, but the average user can hardly distinguish when each is being used or why such a large model has been installed on their hard drive.
Silent download, automatic reinstallation, and hidden options
The real focus of the controversy is not so much that Chrome integrates a local AI model as such, but the way it is distributed and managedIn many cases, the 4GB file appears after the browser has been open for a few minutes, even with only one tab inactive, without the user having consciously activated any artificial intelligence function or seen any confirmation message.
Various testimonies agree that Manually deleting the weights.bin file does not solve the problemIf the user doesn't change Chrome's internal settings, the browser will re-download the template in the background on subsequent startups. In other words, manual deletion temporarily frees up space, but the next session can reuse those 4 GB without any further prompts.
Google says that, since February, it has begun rolling out an option in Chrome's settings to disable AI on the device and delete the local model, thus preventing it from being downloaded or updated again. This option is usually found in the section of Generative AI or system from the settings menu, under a heading similar to "AI on the device" or "On-device AI", and by disabling it the browser should stop managing Gemini Nano and proceed to its removal.
The problem is that this feature isn't available to everyone yet. Several users have reported that, at least in some installations of Chrome for macOS Or in certain regions, the option doesn't appear or appears inconsistently, suggesting that the rollout is being gradual. Meanwhile, specialized media outlets have published guides for manually deleting the file On Windows, macOS, and Linux, it warns that if the internal parameters are not adjusted, the model will be downloaded again.
As an advanced alternative, the use of chrome :/ / flagsThe experimental features panel allows you to disable fields related to local AI, such as "Enables optimization guide on device" or specific flags linked to the Gemini Nano Writer API. These flags allow you to slow down the download or use of certain components, but They are not a user-friendly method for the average user and may have side effects on other browser functions.
Control from Windows 11 and other ways to block downloads
In corporate and professional environments, Microsoft has introduced in Windows 11 An additional way to gain more control over these types of local AI models. This is a policy called GenAILocalFoundationalModelSettings, which allows administrators and advanced users to decide whether the browser can store and run generative models on the device.
In the specific case of Google Chrome, this policy can be configured through the system registry in the path HKLM\SOFTWARE\Policies\Google\ChromeSetting the value GenAILocalFoundationalModelSettings In setting 1, the use of local models is marked as disabled. According to the available information, when this policy is active, the browser It prevents new downloads of the model and may even delete the one that was already installed..
This same logic applies in Microsoft Edgewhich is also incorporating local AI capabilities and recognizes the same policy, albeit with its own configuration branch. For European companies and organizations concerned with data management, bandwidth, and regulatory compliance, these centralized policies offer a more robust path to block installed models without direct user interaction.
In systems such as LinuxSome advanced users have opted for more drastic solutions, such as changing the permissions of the OptGuideOnDeviceModel folder so that Chrome cannot write to it. Using terminal commands that revoke access to the directory (for example, assigning ownership to root and permissions 000), it's possible to prevent the browser from saving the 4 GB to that location again, although this is clearly a solution for highly technical users.
Beyond these measures, experts recommend carefully reviewing the browser's AI settings and, if necessary, evaluating alternatives with clearer transparency policies in the handling of intelligent functions, especially in professional, educational or public administration settings within the European Union.
Privacy, European regulation and the role of consent
The Gemini Nano case in Chrome has sparked a debate that goes far beyond the 4 GB of disk space it occupies. Privacy experts, including [name of expert], have raised concerns. Alexander HanffThey have raised doubts about whether the silent installation of an AI model of this size It respects the principles of transparency and consent set out in European regulations, especially in the ePrivacy Directive and the General Data Protection Regulation (GDPR).
Hanff argues that the automatic download of weights.bin could enter into conflict with Article 5.3 of the ePrivacy Directivewhich requires prior and informed consent before storing or accessing information on the user's device, beyond what is strictly necessary to provide the requested service. In their opinion, several problematic factors combine here: lack of clear notification, absence of a specific opt-in, persistent storage in a hidden directory, and automatic reinstallation if the user tries to delete the file.
The controversy is amplified because Google is not just any player in the European digital ecosystem, but a “gatekeeper” recognized by the Digital Markets Regulation (DMA)Subject to enhanced competition and transparency obligations. Data protection authorities in countries such as Spain (AEPD) or France (CNIL) could be interested in these practices, especially if it is confirmed that the downloading occurred on a large scale within the European Union without a robust consent mechanism.
In addition to the legal dimension, there is a trust component: many users assume that a browser will not install malware on their computer. multi-gigabyte packages without askingThis is especially true when dealing with AI models whose workings are difficult for the general public to understand. In the past, Chrome has distributed additional components, such as cleaning tools in collaboration with third parties, but in those cases, there was generally more communication and control for the end user.
Faced with this scenario, other browsers are trying to differentiate themselves. Firefox, for example, has recently grouped its AI features under a set of “AI controls” These features are accessible from the settings, where it's explained exactly what is being activated and where its future deployment can be blocked from a single panel. Brave, for its part, documents in detail how the models associated with its Leo assistant are downloaded and used, with explicit options to disable or limit their use.
Environmental impact and scale of the massive discharge
Beyond hard drive space or potential regulatory conflicts, there is another, unusual aspect to a controversy of this kind: environmental impactHanff has made rough calculations of what it would mean if this 4GB file were downloaded across the entire Chrome installation base, which numbers over 2.000 million devices worldwide.
If we take that volume of installations as a reference, we would be talking about a potential data transfer of several exabytes, with the consequent energy consumption in data centers and distribution networksTo this must be added the energy associated with storage and processing in users' equipment. According to the estimates cited by the researcher, the massive deployment of this model could be equivalent to between 6.000 and 60.000 tons of CO2 issued, depending on the assumptions used in the calculation.
For an individual user, the download might be limited to a specific peak consumption and a few gigabytes less on their SSD, but on a global scale it becomes one of the largest unsolicited data transfers registered to date. And if you take into account that many users will delete the file upon discovering it and Chrome will download it again, the aggregate impact multiplies.
In a context where artificial intelligence is already under scrutiny for its carbon footprint And the intensive use of resources in training models, the decision to deploy large local models without a prior communication campaign or clear opt-in options, certainly doesn't help improve public perception. Some voices are calling for technology companies to incorporate environmental impact assessments and user-accessible summaries when they implement deployments of this type.
In Europe, where sustainability and energy efficiency policies are a cornerstone of many public and corporate strategies, this discussion could gain traction. It's not unreasonable to think that, in the future, massive deployments of AI in personal devices could be the subject of... additional environmental transparency requirements or periodic reports addressed to regulators.
This whole episode makes it clear that the integration of Gemini Nano in Chrome It's much more than a technical detail: it involves issues of privacy, regulation, user trust, impact on local resources, and sustainability. While Google refines its configuration options and clarifies its position, many users in Spain and the rest of Europe are discovering that, without asking, they are coexisting in their browser with a 4GB AI model that decides for itself when to install, update, or reappear, forcing them to reconsider the extent to which we truly control what happens on our own devices.
