Google trims Gemma 4 for local use as it pushes AI onto laptops and phones

Alphabet’s Google on Friday released new Gemma 4 checkpoints trained with quantization-aware training, or QAT, a method that bakes compression into the training process so models can run with less memory and better speed on laptops, consumer GPUs and other edge devices. The company said the update is designed to make Gemma 4 more efficient for local use rather than only for cloud deployment.

Google said the new release includes checkpoints for the widely used Q4_0 quantization format and for a new mobile-focused format. In that configuration, the company said the Gemma 4 E2B model’s memory footprint falls to 1GB. Google added that QAT is meant to limit the quality loss that can happen when models are compressed after training, because the quantization process is simulated during training instead.

The post lands about two months after Google first introduced Gemma 4, and just days after the company said it had added a 12B version to the family. Google describes Gemma 4 as an open model line built to run and fine-tune efficiently across Android devices, laptop GPUs, workstations and accelerators, with support for longer context windows, multimodal inputs, code generation and agent-like workflows.

Google also said the checkpoints are already integrated with a broad set of developer tools, including, Ollama, LM Studio, LiteRT-LM,, SGLang, vLLM, MLX and Unsloth, and that the weights are available on Hugging Face. That ecosystem support matters because local AI projects often stall if developers cannot easily run the model in the tools they already use.

The release suggests Google is trying to keep its open-model line competitive on efficiency, not just raw capability. That is an inference from the product framing: if the compressed models preserve quality, they can make privacy-sensitive, low-latency and lower-cost AI applications more practical on personal hardware. Google did not disclose pricing, customer-specific terms or a separate commercial rollout in the post.

For Alphabet, the update fits a broader strategy of splitting work between lightweight local models and heavier cloud systems. The likely business value is a wider funnel for developers building consumer apps, coding assistants and edge AI products, while keeping more demanding workloads inside Google’s broader Gemini stack.