Google releases Gemma 3 270M, an ultra-efficient open-source AI model for smartphones

Google DeepMind’s Gemma 3 270M targets developers building AI systems that run directly on devices without relying on cloud infrastructure. Quantized to INT4, the model powered 25 conversation turns on a Pixel 9 Pro using just 0.75 percent of battery. It offers pretrained and instruction-tuned versions, along with Quantization-Aware Training checkpoints for deployment in constrained environments.

The model achieves a 51.2 percent score on IFEval, outperforming similarly sized models and approaching the performance of larger billion-parameter models. Google has made model weights and deployment recipes available through platforms including Hugging Face, Vertex AI, llama.cpp, Gemma.cpp, and JAX.

Gemma 3 270M is aimed at enabling AI applications such as offline assistants, privacy-preserving chatbots, and embedded analytics tools. Its design supports rapid fine-tuning, making it viable for enterprise use cases requiring customization and compliance. The ability to run AI locally reduces network dependency, operational costs, and energy consumption while enabling continuous access in low-connectivity settings.

Google DeepMind stated that Gemma 3 270M represents a step toward specialized, efficient models as an alternative to scaling ever-larger AI architectures. This approach could make AI more accessible for developers and organizations that prioritize control, cost efficiency, and hardware independence.