Google rolls out “hyper-efficient” Gemma 3 270M open AI model

Google has introduced the highly specialized tool Gemma 3 270M, a 270-million parameter model for developers who require task-specific fine-tuning.

According to Google, Gemma 3 270M has a large vocabulary of 256k tokens (small pieces of information used for authentication and authorization), allowing it to handle specific and rare tokens. It also boasts low power consumption using just 0.75% of the battery for 25 conversations in tests, which makes it the most power-efficient Gemma model.

QAT checkpoints are included, allowing INT4 runs with minimal performance degradation.

The model is not designed for complex conversational use cases, although it carefully follows general instructions.

Google released its first light open-source Gemma models, Gemma 3 and Gemma 3 QAT, earlier this year, delivering state-of-the-art performance for single cloud and desktop accelerators.

Gemma 3 270M is built on the same architecture as the rest of the Gemma 3 models and can be easily used for a high-volume, well-defined task, handling both text and images.

The model is accessible on Hugging Face, Ollama, Kaggle, LM Studio, or Docker, with trials on Vertex AI or with popular inference tools like llama.cpp Gemma.cpp, LiteRT, Keras, and MLX.

Google is releasing both pretrained and instruction tuned models.

The company has also announced new multimodal models in the MedGemma collection, used for health AI development.

The first is MedGemma 27B Multimodal, which supports complex multimodal and longitudinal electronic health record interpretation. It’s recommended for tasks that require generating free text or visual question answering.

The second new model is MedSigLIP, an image and text encoder, designed for classification, search, and related tasks.

Both models can be run on a single GPU (or adapted to run on mobile hardware) and are meant to aid in medical research and product development.