Google Chrome has become the browser we use almost without thinking: we open it to work, shop, manage our banking, or watch a show, and we rarely stop to look at what it's doing behind the scenes. That's why it's been so surprising that, when checking their storage, many users have stumbled upon a hidden folder of more than 4 GB occupied by an artificial intelligence model that no one remembers authorizing.
Behind that wasted space lies a file called weights.bin , which is part of an AI model integrated directly into Chrome. This isn't a small, routine update, but rather a multi-gigabyte component that the browser downloads in the background, without a clear dialog box, a visible acceptance checkbox, and, in many cases, without the user even being aware that anything is being installed.
What exactly is Chrome downloading: the OptGuideOnDeviceModel folder and the weights.bin file
The controversy stems from a folder with a rather unintuitive name: OptGuideOnDeviceModel . Inside, there's a file called weights.bin , around 4 GB in size, which doesn't come from any suspicious file extensions or malware. Everything indicates that it's the Gemini Nano model , the scaled-down version of Google's generative AI system designed to run directly on the device.
This local model powers AI features integrated into the browser, such as Google Personal Intelligence , text writing assistance, intelligent suggestion generation, content summaries, and fraud and scam detection on websites. The technical logic is clear: if the processing is done on the user's own computer, some of the information doesn't have to constantly travel to Google's servers.
The problem is that, while it might make sense from an engineering perspective, the way this model is being implemented is practically invisible . Users who have barely opened a tab and left it idle for a few minutes have noticed Chrome starting to download that file without displaying any special notification.
Researchers like privacy expert Alexander Hanff have documented the process using system logs—for example, through kernel logs on macOS—and have found that even new Chrome profiles, without extensions or human interaction , download and unpack the 4GB model in about 15 minutes.
How does Gemini Nano get to your computer and which systems are being affected?
The integration of Gemini Nano into Chrome doesn't work like a traditional download that we consciously accept. In the developer documentation, Google explains that the browser's AI capabilities are designed to be "fluid," and that model management happens automatically in the background . In practice, this means the file might arrive with a routine update or be triggered when an AI feature needs the model for the first time.
Currently, most documented cases point to desktop and laptop computers , especially with macOS and recent versions of Chrome, such as 147. Several reports indicate that, in Spain and the rest of Europe, the behavior is similar: the internal Chrome folders are created, OptGuideOnDeviceModel appears, and shortly after, weights.bin is downloaded and extracted without explicit intervention.
On Windows, some users report finding the file in locations such as %LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel , while on macOS it's typically located within the application support folders in the user profile. There isn't a single universal path, but the pattern is consistent: same folder, same filename, and a size close to 4 GB.
On mobile devices, the situation is different. On many recent Android devices, AI Mode in Chrome is distributed through system components (such as AI Core), so Chrome relies on this existing model. On iOS, file system control is much more restrictive, and in theory, the browser has less leeway for these kinds of large, silent downloads.
It's worth noting that not all users see the update file at the same time. Similar to other Google updates, the rollout appears to be gradual and segmented , meaning the update will arrive in waves to different markets and configurations. However, the trend suggests that it will eventually reach most active Chrome installations.
What does this AI model do inside Chrome and why does it take up so much space?
Gemini Nano isn't simply a chatbot embedded in a sidebar. It's a large-scale language model adapted to run locally , capable of interpreting text, detecting patterns, and generating content. To function, it needs to store millions of parameters, which are, to put it simply, the mathematical "weights" it uses to make decisions.
These parameters are precisely defined in the weights.bin file . Although it's referred to as a "light" version compared to larger models, the practical result is that the file size is around 4 GB . On a PC with a 1 TB SSD, this might go unnoticed, but on laptops with 128 or 256 GB, or on older machines, it represents a very significant loss of space.
Among the functions this model enables are writing assistance within forms, text boxes, and web documents , as well as phrase suggestions, automatic summaries, and contextual translations. It is also mentioned for its use in detecting fraud and scams related to technical support and potentially fraudulent websites, providing an additional layer of security by analyzing the content the user is viewing.
At first glance, the idea of processing this data on the device itself seems positive for privacy: less traffic to the cloud and less dependence on external servers. However, some analyses indicate that not all functions related to Chrome's "AI Mode" are processed locally , a fact that impacts searches and media . Certain tools continue to send keystrokes and content to Google's servers, so the local model coexists with cloud services and does not completely replace them.
This is where some of the discontent arises: many users feel they are giving up disk space and hardware resources without being entirely clear about what is processed locally and what is sent elsewhere , or what concrete benefits they get in return for that silent 4 GB installation.
Silent download: lack of warning, automatic reinstalls, and scattered controls
Beyond the size, what has raised concerns is how Chrome handles the download . Several technical audits have shown that the browser initiates the model transfer in the background, without asking for specific user confirmation and without displaying a message such as "this feature will install a 4 GB file on your computer."
In some documented cases, the sequence has always been the same: a clean installation of Chrome, opening a tab, a few minutes of inactivity , and then the write operations associated with OptGuideOnDeviceModel begin. The download runs even if the user hasn't consciously touched any AI features, reinforcing the feeling that the installation occurs automatically.
Another sticking point is that deleting weights.bin doesn't permanently solve the problem . Users who have tried to free up space by deleting the folder have found that Chrome rebuilds it and re-downloads the model shortly afterward, usually after restarting the browser or following an update. This behavior, described by Hanff with a blunt "Chrome didn't ask, Chrome doesn't show it, Chrome re-downloads it," has resonated with the privacy-conscious community.
Regarding the controls, the situation is far from clear. Some versions of Chrome are starting to include an "AI on device" section in the "System" or advanced settings , which would allow users to disable these features and, in theory, delete the model. In other cases, users have had to resort to the experimental features section at chrome://flags , searching for options like "Enables optimization guide on device" or settings related to on-device AI or Gemini Nano to disable them manually.
The practical consequence is that there is still no single, transparent dashboard for managing all the integrated AI . Meanwhile, some features continue to activate more or less silently, and many people only discover the model exists once it has already consumed several gigabytes of their storage.
Impact on storage, performance, and resource consumption
Chrome has long had a reputation as a RAM-guzzling browser , especially when tabs and extensions accumulate. Now, with the arrival of local AI models, the pressure is also extending to storage, a critical resource in laptops, mini-PCs, and devices with modest SSDs.
A 4GB model might be acceptable in professional environments or on high-capacity machines, but for those already short on space, the situation changes. Suddenly losing several gigabytes without a clear explanation can force users to uninstall applications, move files to the cloud, or forgo updating other programs.
In terms of performance, the presence of a model of this size means that Chrome must load parts of the file into memory when running AI tasks . This can result in increased RAM usage and CPU or GPU load, depending on how Gemini Nano runs on each platform. Devices with limited hardware might experience occasional slowdowns when these functions are activated.
Furthermore, the download and unpacking process itself has its costs: it consumes bandwidth, processor cycles, and SSD write operations . On metered connections or mobile data plans shared with a computer, those 4 GB can represent an entire portion of the monthly allowance, especially in Spain, where many users still have plans with tight limits.
In short, the combination of silent downloads, automatic reinstalls, and resource consumption without any understandable warning turns what could be a technical improvement into a source of frustration for those who see Chrome increasingly taking up space and resources on their computer.
Potential conflicts with GDPR and ePrivacy Directive in Europe
Beyond the technical aspects, the case has sparked all sorts of legal debates in the European Union. Data protection experts point out that the mandatory installation of a 4GB AI model , without specific consent and without clear prior information, could clash with several key European regulations.
On the one hand, Article 5(3) of the ePrivacy Directive (2002/58/EC) establishes that information cannot be stored on, or accessed from, a user's device without their informed consent, except in very specific circumstances. On the other hand, the General Data Protection Regulation (GDPR) requires adherence to principles such as transparency, data minimization, and data protection by design and by default.
From the perspective of researchers like Hanff, downloading a model of this size "just in case" the user ever uses it, without clearly explaining what is being installed or why, could be considered a violation of these principles . It has also been pointed out that practices of this kind fit with what European regulators call dark patterns : design patterns that make it difficult or impossible for users to make informed decisions about their data and their device.
In countries like Spain, where the Spanish Data Protection Agency (AEPD) has already sanctioned major tech companies for a lack of transparency in data processing, this type of behavior comes under scrutiny. To date, no formal action has been announced regarding Gemini Nano in Chrome, but the case fits into ongoing debates about the use of AI and compliance with European regulations.
Added to this is a less visible but relevant aspect: the environmental impact of pushing a 4GB file to hundreds or billions of devices . According to estimates cited by Hanff, a deployment of this scale could result in emissions of between 6.000 and 60.000 tons of CO2 equivalent , depending on the number of devices affected and the energy used in the transfer and processing.
How to tell if you have the 4GB model installed and what you can do
For anyone who suspects that Chrome has suddenly taken up several gigabytes of space, the first step is to check if the file is present. You don't need to be an expert , but it's a good idea to follow some approximate paths depending on your operating system.
En WindowsA frequently mentioned location in various reports is the following (it may vary slightly by profile and version):
%LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel
Inside, you should find a file called weights.binIf its size is around 4 GB, it is almost certainly the local AI model.
On macOS , the folder usually appears in the Chrome application support directories associated with the user. While the exact path may vary depending on the installation, the trick is to search for “OptGuideOnDeviceModel” in Finder within your user folder. Again, if the weights.bin file is several gigabytes in size, you've found the disk hog.
Locating the file is just the beginning. If you simply delete it, Chrome will most likely download it again as soon as it gets the chance , especially if you continue using built-in AI features or if the browser considers them active by default. For more permanent removal, you also need to adjust the settings.
In the most recent versions, some users report the presence of "On-Device AI" options in the System or Privacy and Security section of Chrome's settings. Disabling these settings before deleting the folder would increase the chances that the file won't reappear immediately.
As an additional step, you can go to chrome://flags and search for terms like “optimization guide on device” or references to Gemini Nano, local AI, or models on the device . Disabling these experimental features usually reduces the chances of the browser insisting on reinstalling the model. However, be careful when changing these options, as they affect internal components and may change their name or behavior with each update.
In any case, there's a trade-off: by disabling these features and removing the model, you may lose writing aids, automatic summaries, or other "smart" features . Each user will have to weigh whether they prefer to preserve free space and greater control over their hardware or whether maintaining these local AI capabilities is worthwhile.
In this context, other browsers are taking the opportunity to differentiate themselves. Firefox , for example, has incorporated a dedicated "AI Controls" section in its settings since version 148, where users can block current and future enhancements from a single, easily accessible location. Sections such as on-device AI, translations, and chatbot providers are separated, making it easier to understand what is being enabled and disabled without having to hunt through hidden options in the menu.
The Chrome situation serves as an example of where consumer software is headed: more and more applications are incorporating local AI models to strengthen security, accelerate tasks, or add advanced features. But it also makes it clear that if this transition isn't accompanied by simple explanations, visible controls, and strict adherence to regulations, the average user will find themselves with massive AI models on their computer without really knowing how they got there or what they actually do.
The silent rollout of Gemini Nano in Chrome illustrates the extent to which the race for artificial intelligence is reshaping the relationship between Big Tech and users' devices: our computers are starting to be burdened with multi-gigabyte models that we haven't consciously chosen , and our sense of control is eroding. Until Google offers a more detailed public explanation or a clear, centralized settings panel, the debate will continue to revolve around the same idea: AI can bring useful features to the browser, but not at the cost of colonizing the computer's storage and resources without clear permission.
