Today, we’re adding a new, highly specialized tool to the Gemma 3 toolkit: Gemma 3 270M, a compact, 270-million parameter model.
Google has introduced Gemma 3 270M, a compact AI model designed for hyper-efficient, task-specific applications. With 270 million parameters—170 million for embeddings and 100 million for transformer blocks—it offers a large vocabulary of 256,000 tokens, enabling effective handling of specialized and rare terms. (developers.googleblog.com)
Key Features:
-
Energy Efficiency: Internal tests on a Pixel 9 Pro SoC demonstrated that the INT4-quantized model consumed just 0.75% of the battery after 25 conversations, making it Google’s most power-efficient Gemma model to date. (developers.googleblog.com)
-
Instruction Following: The model is instruction-tuned, capable of following general prompts out of the box, suitable for tasks like data extraction and classification. (developers.googleblog.com)
- Quantization for Production: Quantization-Aware Training (QAT) checkpoints are available, enabling INT4 precision operation with minimal performance degradation, crucial for deployment on resource-constrained devices. (developers.googleblog.com)
Ideal Use Cases:
Gemma 3 270M is well-suited for high-volume, well-defined tasks such as sentiment analysis, entity extraction, query routing, and compliance checks. Its compact size allows for rapid fine-tuning, enabling quick iteration and deployment. Additionally, its energy efficiency and ability to run on-device make it suitable for privacy-sensitive applications. (developers.googleblog.com)
Getting Started:
Developers can access both pretrained and instruction-tuned versions of Gemma 3 270M through platforms like Hugging Face, Ollama, Kaggle, LM Studio, and Docker. The model supports popular inference tools such as llama.cpp, Gemma.cpp, LiteRT, Keras, and MLX, and can be fine-tuned using frameworks like Hugging Face, UnSloth, and JAX. (developers.googleblog.com)
For a practical demonstration, Google showcased a Bedtime Story Generator app powered by Gemma 3 270M, highlighting its capability to generate creative content efficiently. (developers.googleblog.com)
By offering a compact yet powerful model, Gemma 3 270M empowers developers to build efficient, specialized AI applications across various domains.
Discover DeepMind, a world-leading AI research lab by Google. Learn how it’s advancing science, healthcare, and technology through cutting-edge artificial intelligence breakthroughs..
