Google debuts Gemini-2 Nano for edge devices with reduced compute and improved safety

AI · 5 min read

Gemini-2 Nano is Google's response to demand for smaller, safer LLMs that can run with limited power and compute. The model uses a combination of quantization-aware training and conditional sparsity to achieve a compact footprint while retaining strong reasoning and multimodal capabilities tailored for on-device tasks.

Google emphasized safety-by-design: the Nano variant includes built-in filter layers that enforce content guidelines locally and support enterprise policies. The model also includes an audit log API so device manufacturers can surface why particular outputs were generated without exposing raw user data.

Google is partnering with handset makers and AR/VR OEMs to integrate Gemini-2 Nano for private assistants, camera-aware prompts, and local transcription. The company plans a beta SDK with hardware acceleration for later this quarter.