The past few months have been an exciting time for the open model Gemma family. Introducing Gemma 3 and Gemma 3 QAT to deliver cutting-edge performance in a single cloud and desktop accelerator. Next, we announced the full release of Gemma 3n, a mobile-first architecture that brings powerful real-time multimodal AI directly to edge devices. Our goal is to give developers useful tools to build with AI, and as we celebrate together last week with over 200 million downloads, we continue to be amazed by the vibrant Gemmaverse you’re helping create.
Today we are adding a new highly specialized tool to the Gemma 3 toolkit. Gemma 3 270M is a compact 270 million parameter model designed from the ground up for task-specific fine-tuning, with powerful instruction following capabilities and pre-trained text structuring capabilities.
Gemma 3 270M brings powerful command-following capabilities to a space-saving model. It establishes a new level of performance for its size, making advanced AI capabilities more accessible for on-device and research applications, as demonstrated by the IFEval benchmark (testing a model’s ability to follow verifiable instructions).
Core features of Gemma 3 270M
Compact and capable architecture: Our new model has a total of 270 million parameters. 170 million embedding parameters due to large vocabulary size and 100 million for trans blocks. Thanks to its large vocabulary of 256,000 tokens, this model can handle certain rare tokens, making it a powerful base model that can be further fine-tuned for specific domains and languages. Extreme energy efficiency: The main advantage of Gemma 3 270M is its low power consumption. Internal testing of the Pixel 9 Pro SoC showed that the INT4 quantization model used just 0.75% of the battery over 25 conversations, making it the most power-efficient Gemma model. Next instructions: Instruction-tuned models are released with pre-trained checkpoints. Although this model is not designed for complex conversational use cases, it is a powerful model that follows common instructions right out of the box.
In engineering, success is determined by efficiency, not just raw power. Never use a sledgehammer to hang a picture frame. The same principles apply to building with AI.
The Gemma 3 270M embodies this ‘right tool for the job’ philosophy. A high-quality basic model that follows instructions faithfully right out of the box, its true power is unlocked with fine-tuning. Specialization allows you to perform tasks such as text classification and data extraction with incredible accuracy, speed, and cost efficiency. By starting with a compact and functional model, you can build a production system that is lean, fast, and significantly cheaper to operate.
A blueprint for real-world success
The power of this approach is already producing surprising results in the real world. A perfect example is our Adaptive ML work with SK Telecom. Faced with the challenge of moderating sensitive multilingual content, they chose to specialize. Instead of using a large generic model, Adaptive ML fine-tuned the Gemma 3 4B model. The results were amazing. The specialized Gemma model not only met, but exceeded the performance of a much larger proprietary model for its particular task.
The Gemma 3 270M is designed to allow developers to take this approach further and achieve even greater efficiency for well-defined tasks. This is the perfect starting point for creating a fleet of small, specialized models, each expert at its own task.
However, this power of specialization is not only useful for corporate tasks. It also enables powerful creative applications. For example, check out this Bedtime Story Generator web app:
Gemma 3 270M is used to run the Bedtime Story Generator web app using Transformers.js. The model’s size and performance make it suitable for offline, web-based, and creative tasks. (Credit: Joshua from the Hugging Face team (@xenovacom on X))
If you choose Gemma 3 270M
Gemma 3 270M inherits the advanced architecture and robust pre-training of the Gemma 3 collection, providing a solid foundation for custom applications.
The best choices are:
There are a large number of well-defined tasks. Ideal for features such as sentiment analysis, entity extraction, query routing, unstructured to structured text processing, creative writing, and compliance checking. Every millisecond and microcent needs to be counted. Significantly reduce or eliminate inference costs in production and provide faster responses to users. The fine-tuned 270M model can run on lightweight, inexpensive infrastructure or directly on your device. You need to iterate and deploy quickly. The Gemma 3 270M’s small size allows for rapid fine-tuning experiments, allowing you to find the best configuration for your use case in hours instead of days. User privacy must be ensured. This model runs completely on-device, allowing you to build applications that process sensitive information without sending data to the cloud. A fleet of specialized task models is required. Build and deploy multiple custom models, each professionally trained for a different task, without compromising your budget.
Start fine-tuning
We want to make it as easy as possible to turn your Gemma 3 270M into your own custom solution. Built on the same architecture as other Gemma 3 models, it comes with recipes and tools to get you started quickly. A complete guide to fine-tuning using the Gemma 3 270M can be found as part of the Gemma documentation.
Gemmaverse is built on the idea that innovation comes at all scales. Gemma 3 270M enables developers to build smarter, faster, and more efficient AI solutions. I can’t wait to see the special models you create.
