gemma-3-270m

Gemma 3 270M is a compact, text-only open-weight language model from Google DeepMind, released in August 2025 as the smallest member of the Gemma 3 family. With roughly 270 million parameters—about 170M in embeddings (256K vocabulary) and 100M in transformer blocks—it is designed for task-specific fine-tuning rather than general-purpose chat. It supports a 32K token context window and was pretrained on 6 trillion tokens spanning web documents, code, and mathematics, with a knowledge cutoff of August 2024.

The model ships in both pretrained and instruction-tuned (IT) variants, plus Quantization-Aware Trained (QAT) INT4 checkpoints for production deployment. On official benchmarks, the instruction-tuned variant scores 51.2 on IFEval and 26.7 on BIG-Bench Hard, establishing strong instruction-following for its size class. Google positions it for high-volume, well-defined tasks such as text classification, entity extraction, sentiment analysis, query routing, and structured data extraction.

Gemma 3 270M is optimized for extreme energy efficiency and privacy-preserving on-device use. Internal tests on Pixel 9 Pro SoC show INT4-quantized inference consuming just 0.75% battery for 25 conversations. It is available under the Gemma Terms of Use via Hugging Face, Kaggle, Ollama, Vertex AI, and supports fine-tuning with Hugging Face, UnSloth, JAX, and Keras.