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

# Introducing Gemma 3 270M: The compact model for hyper-efficient AI Source: https://developers.googleblog.com/en/introducing-gemma-3-270m/ The last few months have been an exciting time for the Gemma family of open models. We introduced [Gemma 3](https://blog.google/technology/developers/gemma-3/) and [Gemma 3 QAT](https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/), delivering state-of-the-art performance for single cloud and desktop accelerators. Then, we announced the full release of [Gemma 3n](https://developers.googleblog.com/en/introducing-gemma-3n/), a mobile-first architecture bringing powerful, real-time multimodal AI directly to edge devices. Our goal has been to provide useful tools for developers to build with AI, and we continue to be [amazed](https://www.youtube.com/watch?v=Fx6IuEggeac) by the vibrant [Gemmaverse](https://deepmind.google/models/gemma/gemmaverse/) you are helping create, celebrating together as downloads surpassed 200 million last week. Today, we're adding a new, highly specialized tool to the Gemma 3 toolkit: [Gemma 3 270M](https://ai.google.dev/gemma/docs/core/huggingface_text_full_finetune), a compact, 270-million parameter model designed from the ground up for task-specific fine-tuning with strong instruction-following and text structuring capabilities already trained in. ## Gemma 3 270M Gemma 3 270M brings strong instruction-following capabilities to a small-footprint model. As shown by the IFEval benchmark (which tests a model's ability to follow verifiable instructions), it establishes a new level of performance for its size, making sophisticated AI capabilities more accessible for on-device and research applications. ## Core capabilities of Gemma 3 270M - **Compact and capable architecture:** Our new model has a total of 270 million parameters: 170 million embedding parameters due to a large vocabulary size and 100 million for our transformer blocks. Thanks to the large vocabulary of 256k tokens, the model can handle specific and rare tokens, making it a strong base model to be further fine-tuned in specific domains and languages. - **Extreme energy efficiency:** A key advantage of Gemma 3 270M is its low power consumption. Internal tests on a Pixel 9 Pro SoC show the INT4-quantized model used just 0.75% of the battery for 25 conversations, making it our most power-efficient Gemma model. - **Instruction following:** An instruction-tuned model is released alongside a pre-trained checkpoint. While this model is not designed for complex conversational use cases, it's a strong model that follows general instructions right out of the box. - **Production-ready quantization:** [Quantization-Aware Trained](https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-state-of-the-art-ai-to-consumer-gpus/) (QAT) [checkpoints are available](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d), enabling you to run the models at INT4 precision with minimal performance degradation, which is essential for deploying on resource-constrained devices. In engineering, success is defined by efficiency, not just raw power. You wouldn't use a sledgehammer to hang a picture frame. The same principle applies to building with AI. Gemma 3 270M embodies this "right tool for the job" philosophy. It's a high-quality foundation model that follows instructions well out of the box, and its true power is unlocked through fine-tuning. Once specialized, it can execute tasks like text classification and data extraction with remarkable accuracy, speed, and cost-effectiveness. By starting with a compact, capable model, you can build production systems that are lean, fast, and dramatically cheaper to operate. ## A real-world blueprint for success The power of this approach has already delivered incredible results in the real world. A perfect example is [the work done by Adaptive ML with SK Telecom](https://deepmind.google/models/gemma/gemmaverse/adaptiveml/). Facing the challenge of nuanced, multilingual content moderation, they chose to specialize. Instead of using a massive, general-purpose model, Adaptive ML fine-tuned a Gemma 3 4B model. The results were stunning: the specialized Gemma model not only met but exceeded the performance of much larger proprietary models on its specific task. Gemma 3 270M is designed to let developers take this approach even further, unlocking even greater efficiency for well-defined tasks. It's the perfect starting point for creating a fleet of small, specialized models, each an expert at its own task. But this power of specialization isn't just for enterprise tasks; it also enables powerful creative applications. For example, check out [this Bedtime Story Generator web app](https://huggingface.co/spaces/webml-community/bedtime-story-generator): Gemma 3 270M used to power a Bedtime Story Generator web app using Transformers.js. The model's size and performance make it suitable for offline, web-based, creative tasks. (Credit: Joshua (@xenovacom on X) from the Hugging Face team) ## When to 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 your custom applications. Here's when it's the perfect choice: - **You have a high-volume, well-defined task.** Ideal for functions like sentiment analysis, entity extraction, query routing, unstructured to structured text processing, creative writing, and compliance checks. - **You need to make every millisecond and micro-cent count.** Drastically reduce, or eliminate, your inference costs in production and deliver faster responses to your users. A fine-tuned 270M model can run on lightweight, inexpensive infrastructure or directly on-device. - **You need to iterate and deploy quickly.** The small size of Gemma 3 270M allows for rapid fine-tuning experiments, helping you find the perfect configuration for your use case in hours, not days. - **You need to ensure user privacy.** Because the model can run entirely on-device, you can build applications that handle sensitive information without ever sending data to the cloud. - **You want a fleet of specialized task models.** Build and deploy multiple custom models, each expertly trained for a different task, without breaking your budget. ## Get started with fine-tuning We want to make it as easy as possible to turn 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 started quickly. You can find our guide on [full fine-tuning](https://ai.google.dev/gemma/docs/core/huggingface_text_full_finetune) using Gemma 3 270M as part of the Gemma docs. - **Download the model:** Get the Gemma 3 270M models from [Hugging Face](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d), [Ollama](https://ollama.com/library/gemma3), [Kaggle](https://www.kaggle.com/models/google/gemma-3), [LM Studio](https://lmstudio.ai/models/google/gemma-3-270m), or [Docker](https://hub.docker.com/r/ai/gemma3). We are releasing both pretrained and instruction tuned models. - **Try the model:** Try the models on [Vertex AI](https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/gemma3) or with popular inference tools like [llama.cpp](https://huggingface.co/collections/ggml-org/gemma-3-270m-689e0105d56462786413d7fc), [Gemma.cpp](https://www.kaggle.com/models/google/gemma-3/gemmaCpp), [LiteRT](https://huggingface.co/litert-community/gemma-3-270m-it), [Keras](https://www.kaggle.com/models/keras/gemma3), and [MLX](https://huggingface.co/collections/mlx-community/gemma-3-270m-689e1de307ccaeec5ba22ec9). - **Start fine-tuning:** Use your favorite tools, including [Hugging Face](http://ai.google.dev/gemma/docs/core/huggingface_text_full_finetune), [UnSloth](https://docs.unsloth.ai/basics/gemma-3-how-to-run-and-fine-tune), and [JAX](https://gemma-llm.readthedocs.io/en/latest/colab_finetuning.html). - **Deploy your solution:** Once fine-tuned, you can deploy your specialized model anywhere, from [your own local environment](http://localhost:8080/) to [Google Cloud Run](https://cloud.google.com/run/docs/run-gemma-on-cloud-run). The Gemmaverse is built on the idea that innovation comes in all sizes. With Gemma 3 270M, we're empowering developers to build smarter, faster, and more efficient AI solutions. We can't wait to see the specialized models you create.