Google has introduced FunctionGemma, a compact model with 270 million parameters. Its goal is to eliminate the main problem of modern programs: the low reliability of artificial intelligence when working locally on gadgets, rather than in the cloud.
Google has introduced FunctionGemma, a compact model with 270 million parameters. Its goal is to eliminate the main problem of modern programs: the low reliability of artificial intelligence when working locally on gadgets, rather than in the cloud.
Unlike universal chatbots, FunctionGemma is designed for one critical function - converting user commands in natural language into structured code that applications and devices can actually execute, all without connecting to the cloud, writes VentureBeat.
As the industry continues to strive for trillion-parameter scale in the cloud, FunctionGemma is betting on small language models (SLMs) running locally on phones, browsers, and IoT devices.
FunctionGemma provides development teams with a new basic architectural element — a confidential “router” that performs complex logical operations locally and with near-zero latency.
FunctionGemma is now available for download on Hugging Face and Kaggle . You can also download the Google AI Edge Gallery app from the Google Play Store.
According to Google Mobile Actions’ internal evaluation, the typical small model suffers from reliability issues, showing a baseline accuracy of only 58% for function call tasks. However, after fine-tuning for this specific purpose, FunctionGemma’s accuracy jumped to 85%. Now, this tiny model performs as effectively as much larger and more complex neural networks.
This allows the model to process more than just simple on/off switches; it is capable of analyzing complex arguments, such as determining specific grid coordinates to control game mechanics or detailed logic.
The model has 270 million parameters and is trained on the 6 trillion tokens of the Mobile Actions dataset, which will help developers train their own agents. It is also compatible with the Hugging Face Transformers, Keras, Unsloth, and NVIDIA NeMo libraries.
According to the company, the model has the following key advantages:
