Google releases Gemma 3 270M, a small, high-performance AI model that can run on low-spec devices such as smartphones, and is energy-efficient and can be used as a base model for AI specialized in specific fields.

Google released the open model ' Gemma 3 270M ' on Thursday, August 14, 2025. Gemma 3 270M is a relatively small model with 270 million parameters, but it has high performance.

Introducing Gemma 3 270M: The compact model for hyper-efficient AI - Google Developers Blog

Introducing Gemma 3 270M! 🚀 It sets a new standard for instruction-following in compact models, while being extremely efficient for specialized tasks. https://t.co/kC9OOPwzVi

— Google AI Developers (@googleaidevs) August 14, 2025

The Gemma 3 270M is a compact model released as part of the Gemma 3 series . It has 270 million parameters, of which 100 million are Transformer blocks responsible for data processing, and the remaining 170 million are embedded parameters containing a huge vocabulary. The vocabulary is 256,000 tokens, making it possible to handle rare vocabularies. Furthermore, it can be fine-tuned to create AI models specialized for specific domains.

The Gemma 3 270M was trained with a focus on command-following and text structuring capabilities. The graph below shows the number of model parameters on the horizontal axis and the score of the IFEval benchmark test, which measures command-following ability, on the vertical axis. The graph shows that the Gemma 3 270M has significantly higher command-following ability than models of a similar size.

The Gemma 3 270M is released in two models: the standard model 'gemma-3-270m' and the quantized version 'gemma-3-270m-it.' The latter model is capable of INT4 quantization while minimizing performance degradation, allowing it to run at high speed even on low-spec devices. The Gemma 3 270M also boasts power-saving performance. In tests using the Pixel 9 Pro SoC, the Gemma 3 270M's INT4 quantized model consumed only 0.75% of the battery even after 25 conversations.

The Gemma 3 270M is also designed for developers to fine-tune it, allowing them to build smaller, more powerful models for specific applications. According to Google, these smaller, fine-tuned models can sometimes outperform larger models.

gemma-3-270m and gemma-3-270m-it are distributed by Hugging Face. Various AI runtime environments such as Ollama and LM Studio also support running Gemma 3 270M.

Gemma 3 Release - a google Collection
https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d