Chrome's Gemini Nano AI Model Consumes 4GB of Disk Space
According to a report from The Verge, Google Chrome may be silently using up to 4GB of local disk space for its on-device AI features. The culprit is a file named weights.bin located in the OptGuideOnDeviceModel directory under Chrome's data folder. This file contains the training parameters for Google's Gemini Nano model, which powers features like scam detection, writing assistance, autofill, and suggestion tools directly on your machine without cloud calls.
How to Check and Reclaim Storage
To see if the file is present on your system, navigate to your Chrome user data folder. The exact path varies by OS, but typically looks like:
- Windows:
%LOCALAPPDATA%\Google\Chrome\User Data\OptGuideOnDeviceModel - macOS:
~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel - Linux:
~/.config/google-chrome/OptGuideOnDeviceModel
If you delete the weights.bin file manually but keep AI features enabled, Chrome will re-download it on the next update. The correct way to free up the space permanently is to head to Settings > System and toggle off the On-Device AI option. This removes the model and disables AI features that rely on it.
Google's Statement and Caveats
Google spokesperson Scott Westover clarified: “We’ve offered Gemini Nano for Chrome since 2024 as a lightweight, on-device model. It powers important security capabilities like scam detection and developer APIs without sending your data to the cloud. ... The model will automatically uninstall if the device is low on resources.”
As of February 2026, Chrome provides a user-facing toggle to disable and remove the model. However, the 4GB size is only documented in a lengthy help center article, not at the point of enabling the features. Google notes that “Gemini Nano’s exact size may vary as the browser updates the model.”
If you rely on Chrome's AI features but are tight on storage, there is currently no option to switch to a cloud-based model; the only workaround is to disable on-device AI entirely.
📖 Read the full source: HN AI Agents
👀 See Also
Two new models appear on OpenRouter, possibly DeepSeek V4 variants
Two new models named healer-alpha and hunter-alpha have appeared on OpenRouter, with specifications matching leaked details about DeepSeek V4. Initial testing shows both models perform well in roleplay scenarios with no message filtering and faster token generation than GLM 5.0.
Benchmarking the Latest AI Models: The Rise of Extreme Models
A detailed benchmarking of 40 new AI models reveals a split market with 'God Mode' and 'Flash Mode' leading the way. Mid-range models are now considered obsolete.
OpenClaw April Updates: A Month of Breaking Changes and Eroded Trust
OpenClaw's April updates show a pattern: new features and fixes shipped alongside critical bugs. Postinstall scripts deleting files, security holes, and broken skills erode confidence.
KV Cache Architecture Evolution: From GPT-2 to Mamba
Analysis of KV cache memory costs shows GPT-2 used 300 KiB/token, Llama 3 reduced it to 128 KiB/token with grouped-query attention, and DeepSeek V3 achieved 68.6 KiB/token with multi-head latent attention. Mamba/SSMs eliminate KV cache entirely with fixed-size hidden states.