Google unveiled Gemma 4 QAT, a version of its Gemma 4 model family designed to sharply reduce memory use so large AI models can run more easily on smartphones and standard laptops. It applies quantization-aware training, simulating quantization during training to limit quality loss. Google said the approach can preserve response quality while lowering memory needs. The models are free under the Apache License 2.0 and officially support local runtimes such as llama.cpp, Ollama and LM Studio.

A newly disclosed web tracking technique can analyse SSD activity inside a browser to infer other websites a user is viewing and even apps running elsewhere. Researchers presented a contention-based side-channel method called Frost that repeatedly accesses large files in the browser’s Origin Private File System and measures tiny SSD I/O latency differences. They said the data can be analysed with a pre-trained convolutional neural network model. The method faces constraints, including the need for very large files and storage placement on the same SSD.