Google launches Gemma 3 270M: a compact fine-tuning AI model optimized for efficiency

Google has introduced Gemma 3 270M, a new addition to the Gemma 3 toolkit. This compact model features 270 million parameters and is engineered for task-specific fine-tuning. It provides robust instruction-following and text structuring capabilities out of the box.

The model leverages a large vocabulary of 256,000 tokens and 170 million embedding parameters, allowing it to handle uncommon and domain-specific tokens across languages. This makes Gemma 3 270M highly adaptable as a foundation model for further customization. Internal tests highlight substantial energy savings: the INT4-quantized version used only 0.75% battery during 25 conversations on a Pixel 9 Pro system-on-chip, making it Google's most power-efficient Gemma model to date.

Alongside a pre-trained checkpoint, Google has released an instruction-tuned variant that can perform general instruction-following tasks immediately. To support deployment on resource-constrained devices, Quantization-Aware Trained checkpoints allow the model to run at INT4 precision with minimal loss in performance. While Gemma 3 270M is not intended for complex multi-turn dialogues, it responds well to most instructions and is especially effective when fine-tuned for text classification and data extraction. The model is available for download through Hugging Face, Ollama, Kaggle, LM Studio, and Docker, and is also available for testing on Vertex AI.