InfoQ

<h1>Google Expands Gemma 3 with New Multilingual and Function-Calling Models</h1>
<p>Google has made significant advancements in its machine learning offerings with the release of two new model suites: TranslateGemma and FunctionGemma. These models are developed under the overarching Gemma 3 architecture, catering to a wide range of translation and functional calling applications.</p>
<h2>TranslateGemma: Advanced Multilingual Capabilities</h2>
<p>TranslateGemma is an open-source suite designed for multilingual translation tasks, built on Gemma 3 architecture with variants of 4 billion, 12 billion, and 27 billion parameters. This model suite supports translation across 55 languages, achieving notable parameter efficiency by combining supervised fine-tuning with reinforcement learning. Remarkably, the 12 billion parameter model repeatedly outperforms the 27 billion baseline, as demonstrated on the WMT24++ benchmarks.</p>
<p>The models are engineered for flexible deployment, suitable for anything from mobile devices to more powerful cloud accelerators. TranslateGemma retains multimodal capabilities, allowing it to translate text embedded within images effectively. Additionally, the training encompasses nearly 500 extra language pairs, aiding community-driven language research.</p>
<h2>FunctionGemma: Empowering Mobile and Edge Devices</h2>
<p>On the other hand, FunctionGemma is a lighter variant with 270 million parameters, specifically optimized for function calling tasks on mobile and edge devices. This model translates natural language into structured API calls, facilitating on-device AI agents in executing actions such as setting reminders or adjusting system settings.</p>
<p>FunctionGemma is noticeably efficient, operating seamlessly on devices with limited resources, and notably improving mobile action accuracy from 58% to an impressive 85% through fine-tuning. It supports unified action and chat modes, integrating smoothly with platforms like Hugging Face, Keras, and vLLM.</p>
<p>To demonstrate its practicality, Google provides demos such as Mobile Actions, TinyGarden, and Physics Playground, alongside datasets and notebooks for user customization.</p>
<p>These new releases underscore Google’s commitment to advancing AI accessibility and functionality across varying computing environments. By balancing complexity and efficiency, Google positions these models as powerful tools for developers and researchers worldwide.</p>


Collections

Google released two new model suites under Gemma 3: TranslateGemma for multilingual translation across 55 languages with variants up to 27B parameters, and FunctionGemma, a lightweight 270M parameter model for function calling on mobile and edge devices. TranslateGemma achieves strong parameter efficiency through supervised fine-tuning and reinforcement learning, with the 12B model outperforming the 27B baseline. FunctionGemma translates natural language to structured API calls and improves mobile action accuracy from 58% to 85% through fine-tuning, supporting deployment on resource-constrained devices.

Google Expands Gemma 3 with New Multilingual and Function-Calling Models

<p>This is a smart direction from Google.
Most translation models get bigger to get better, but here the 12B beating 27B shows how much good fine-tuning + RL can matter.
Also, FunctionGemma at 270M for on-device actions is very interesting for anyone building mobile AI agents without cloud dependency.
Feels like a real step toward practical edge AI, not just bigger models.</p>
