Google has announced the Gemma 3 270M. It is a compact 270 million parameter model aimed at task-specific fine-tuning and efficient on-device deployment. The release expands the Gemma Open Model family, which surpasses 200 million downloads, following the launch of Gemma 3, Gemma 3 Qat and Gemma 3N.
The company said the model is built for instructional and text structuring, but is also energy efficient. Internal testing on the Pixel 9 Pro SoC showed an Int4 Quantised version that consumes 0.75% of the battery in 25 conversations.
“The Gemma 3 270m embodies a tool that is suitable for job philosophy,” the company said in a blog post. “This is a high quality basic model, it's a clean and follow instructions out of the box, and its true power is locked through fine tuning.”
The model includes 170 million embedded parameters to support a large 256k vocabulary and 100 million transformer parameters, suitable for rare tokens and domain-specific fine-tuning. Quantisation-Aware Training (QAT) checkpoints are also available, allowing for INT4 accuracy deployment with minimal performance impact.
Google said the model is not designed for complex conversation use cases, but can be specialized for applications such as text classification, entity extraction, compliance checks, query routing and creative writing. Developers can also run the device completely on Gemma 3 270m to address privacy-sensitive use cases.
The company pointed to the work of Adaptive ML with SK Telecom as an example of the effectiveness of specialization. By fine-tuning the Gemma 3 4B model for multilingual content moderation, the adaptive ML has achieved performance beyond its much larger proprietary model.
The Gemma 3 270M is already being used in creative projects, such as the Bedtime Story Generator Web App developed with Transformers.js.
Google releases both assumption and instruction tuning checkpoints with Hug, Orama, Kaggle, LM Studio and Docker. This model can be tested with vertex AI or run with inference tools such as llama.cpp, gemma.cpp, litert, keras, mlx. Tweaks are supported by hugging faces, sloths and JAX with deployment options from local environments to Google Cloud Run.
“Gemmaverse is based on the idea that innovation can come from all sizes,” the company said. “With the Gemma 3 270m, developers can now build smarter, faster, and more efficient AI solutions.”
