The last few months have been an exciting time for the Gemma family of open models. we presented Gemma 3 AND Gemma 3QATproviding cutting-edge performance for single cloud accelerators and desktops. Then we announced the full release Gemma 3nmobile-centric architecture, delivering powerful real-time multimodal AI directly to edge devices. Our goal was to provide developers with useful tools to build AI-powered applications, and we still do surprised through the vibrant Gemmaverse you help create as you celebrate the fact that downloads surpassed 200 million last week.
Today we are adding a new, highly specialized tool to the Gemma 3 toolkit: Gemma 3 270Ma compact 270 million parameter model designed from the ground up for task-specific tuning, with advanced instruction execution and text structuring features you've already trained in.
Gemma 3 270M provides extensive instructions in a small model. As demonstrated by the IFEval benchmark (which tests a model's ability to follow testable instructions), it sets a new level of performance for its size, making advanced AI features more accessible to on-device and research applications.
Basic capabilities of the Gemma 3 270M
- Compact and efficient architecture: Our new model has a total of 270 million parameters: 170 million embedding parameters due to the large vocabulary size and 100 million for our transformer blocks. Thanks to the large vocabulary of 256,000 tokens model can handle specific and rare tokens, making it a strong base model that can be further tuned for specific domains and languages.
- Extreme energy efficiency: The key advantage of Gemma 3 270M is low power consumption. Internal testing of the Pixel 9 Pro SoC showed that the INT4 quantized model used just 0.75% of its battery over 25 calls, making it our most power-efficient Gemma model yet.
- Instructions as follows: The instruction-aligned model is released with a pre-trained checkpoint. Although this model is not intended for complex conversational applications, it is a powerful model that follows general instructions straight away.
In engineering, success is defined by performance, not just sheer power. You wouldn't use a hammer to hang a picture frame. The same principle applies to building with artificial intelligence.
The Gemma 3 270M embodies the “right tool for the job” philosophy. This is a high-quality entry-level model that follows the instructions straight out of the box, and its true power is unlocked through fine tuning. Once specialized, it can perform tasks such as text classification and data extraction with extreme accuracy, speed and cost-effectiveness. Starting with a compact, efficient model, you can build production systems that are cost-effective, fast and much cheaper to operate.
A real plan for success
The power of this approach has already produced amazing results in the real world. A perfect example is the work Adaptive ML has done with SK Telecom. Faced with the challenge of nuanced, multilingual content moderation, they decided to specialize. Instead of using a bulky general-purpose model, Adaptive ML has improved the Gemma 3 4B. The results were stunning: Gemma's specialized model not only matched but even exceeded the performance of much larger, proprietary models for a specific task.
The Gemma 3 270M is designed to enable developers to take even greater advantage of this approach, unlocking even greater performance for well-defined tasks. This is the perfect starting point for creating a fleet of small, specialized models, each of which is an expert at its task.
But this power of specialization doesn't just apply to corporate tasks; it also enables advanced creative applications. For example, check this Bedtime Story Generator web app: :
Gemma 3 270M used to run the Bedtime Story Generator web app using Transformers.js. The model's size and performance make it suitable for offline, online creative tasks. (Source: Joshua (@xenovacom on X) from Hugging Face)
When to choose Gemma 3 270M
The Gemma 3 270M inherits the advanced architecture and solid initial training from the Gemma 3 collection, providing a solid foundation for custom applications.
Here's when it's the perfect choice:
- You have a broad and well-defined task. Perfect for functions such as sentiment analysis, entity extraction, querying, unstructured to structured text processing, creative writing and compliance checking.
- You have to count every millisecond and microcent. Drastically reduce or eliminate inference costs in production and deliver faster responses to users. The refined 270M model can run on lightweight, low-cost infrastructure or directly on-device.
- You need to iterate and deploy quickly. The Gemma 3 270M's small size allows for quick tuning experiments, helping you find the perfect configuration for your use case in hours, not days.
- You must ensure user privacy. Because the model can run entirely on the device, you can create applications that handle sensitive information without having to send data to the cloud.
- You need a fleet of specialized task models. Create and deploy multiple custom models, each expertly trained for a different task, without breaking your budget.
Start with tuning
We want to make it as easy as possible for you to transform the Gemma 3 270M into your own custom solution. It's built on the same architecture as the rest of the Gemma 3 models, with recipes and tools to get you up and running quickly. You can find our guide at full tuning using Gemma 3 270M as part of Gemma's documentation.
The Gemmaverse is based on the idea that innovation comes in different sizes. With Gemma 3 270M, we enable developers to create smarter, faster and more efficient AI solutions. We can't wait to see the specialized models you create.
