Google has announced a small version of its open-source Gemma model designed to run on local devices. The company says the new Gemma 3 270M can be quickly configured and maintains high performance despite its small size.
Google has announced a small version of its open-source Gemma model designed to run on local devices. The company says the new Gemma 3 270M can be quickly configured and maintains high performance despite its small size.
Google released its first open-source Gemma 3 models earlier this spring, which contain anywhere from 1 billion to 27 billion parameters. The new Gemma 3 has just 270 million parameters and can run on devices like smartphones or even entirely in a web browser, Ars Technica reports .
Running an AI model locally has numerous benefits, including improved privacy and lower latency. The Gemma 3 270M was designed to do just that. When tested on the Pixel 9 Pro, the new Gemma was able to run 25 calls on the Tensor G4 chip and used only 0.75% of the device’s battery. This makes it the most efficient Gemma model to date.
Developers shouldn't expect the same level of performance as a model with several billion parameters, but the Gemma 3 270M has its advantages. Google used the IFEval benchmark, which tests a model's ability to execute instructions, to show that the new model punches above its weight. The Gemma 3 270M scored 51.2% on this test, which is higher than other lightweight models with more parameters. The new Gemma, as expected, falls short of models with more than a billion parameters, such as Llama 3.2, but it's closer than you might think with only a fraction of the parameters.
Google says the Gemma 3 270M performs well out of the box, but expects developers to fine-tune the model for their specific use cases. The small number of parameters makes this process fast and inexpensive. Google is eyeing the new Gemma for tasks like text classification and data analysis, which it can perform quickly and without large computational demands.
Gemma 3 270M is open, but not to be confused with "open source". However, the situation is similar in many ways. The new Gemma model is free to download, and its weights are available. There is no separate commercial license agreement, so developers are free to modify, publish, and use derivatives of Gemma 3 270M in their tools.
However, anyone using Gemma models is required to comply with terms of use, which prohibit modifying the models to produce harmful results or intentionally violate privacy policies. Developers are also responsible for detailing modifications and providing a copy of the terms of use for any derivative versions that inherit Google's own license.
Gemma 3 270M is available on platforms like Hugging Face and Kaggle in pre-trained and instruction-driven versions. It is also available on Google Vertex AI for testing. The company also demonstrated the new model’s capabilities with a full-fledged story generator running in the browser based on Transformer.js.
