Google has introduced Gemma 3 270M, a compact large language model with 270 million parameters, optimized for on-device deployment with only 550MB of memory. Ideal for rapid model iteration and high-volume tasks, it aims to enhance efficiency despite common issues like hallucinations and copyright concerns. Released alongside earlier versions in February 2024, it offers pre-trained models without source data. Google claims this model outperforms competitors like SmollLM2-360M-Instruct on the IFEval benchmark while prioritizing energy efficiency—showing minimal battery drain on devices like the Pixel 9 Pro. With a training dataset of six trillion tokens, it maintains a similar breadth to larger sibling models. Gemma 3 270M is free but comes with usage restrictions, prohibiting violations of rights, dangerous content, and attempts to circumvent safety filters. Users can access it through platforms like Hugging Face and Kaggle, along with a guide for fine-tuning.
Source link
Share
Read more