Key Points- Google adds Quantization-Aware Training to Gemma 4, cutting memory use.

• New Q4_0 and mobile optimized checkpoints let the models run on everyday devices.
• Smaller models open the door for local AI on ChromeOS and consumer GPUs.

Since releasing Gemma 4 two months ago, Google has kept expanding its capabilities. First, Multi‑Token Prediction was added to speed up inference, and a few days ago a 12B model appeared, filling the space between the earlier E4B and 26B versions. Gemma 4 12B model

Today the team announced new checkpoints that use Quantization-Aware Training, a method that trains the model while simulating the later compression step. This approach reduces quality loss compared to standard Post‑Training Quantization. The release includes Quantization-Aware Training Q4_0 checkpoints for the popular format.

For edge devices Google introduced a special mobile specialized quantization schema that targets low‑memory environments. With this mobile format the E2B version now fits into just 1GB of RAM, dramatically lowering the storage and VRAM needed.

Quantization remains a key way to run large models on consumer hardware, improving decode speed while keeping quality high.

The reduced memory footprint means Chromebook users and other edge platforms can load and run these models directly, without needing a powerful cloud server. This makes AI features in ChromeOS more responsive and opens new possibilities for offline apps.

With these optimizations, developers can experiment locally, test models on modest laptops, and deliver faster, privacy‑focused experiences to everyday users.

Anyone with a modest laptop can now run Gemma 4 locally and see the difference.

Read the rest of the article

You can also check out our list of the Best Instagram Extensions, Best Pinterest Exensions & the Best AI Extensions.