From Gemma 3 270M to FunctionGemma, How Google AI Is Building a Collaborative Task Calling Expert for Edge Workloads

Google has released FunctionGemma, a special version of the Gemma 3 270M model that is specially trained for calling and designed to act as an edge agent that maps natural language to actionable API actions.

But, What is FunctionGemma?

FunctionGemma is a text-only 270M converter based on the Gemma 3 270M. It maintains the same structure as Gemma 3 and is released as an open model under the Gemma license, but the training objective and format of the dialogue is dedicated to the performance of hitting instead of free-form dialogue.

The model is intended to be fine-tuned for specific calling functions. It is not positioned as a general chat assistant. The main design goal is to translate user instructions and tool descriptions into structured function calls, and optionally summarize user tool responses.

From an interface perspective, FunctionGemma is presented as a general causal language model. Input and output are text sequences, with an input context of 32K tokens and an output budget of up to 32K tokens per request, divided by the length of the input.

Construction data and training

The model uses the Gemma 3 transformer architecture and the same 270M parameter scale as the Gemma 3 270M. The training and runtime stack reuses research and infrastructure used for Gemini, including JAX and ML Pathways on large TPU clusters.

FunctionGemma uses Gemma's 256K vocabulary, optimized for JSON structures and multilingual text. This improves the efficiency of tokens in task schemas and tool responses and reduces the length of edge execution sequences when latency and memory are tight.

The model is trained on 6T tokens, with a data cutoff of August 2024. The data set focuses on two main categories:

• community tool and API definitions
• interoperability of input tools, function calls, function responses and natural language follow-up messages that summarize output or clarification

This training signal teaches both syntax, function calling and how to format arguments, as well as purpose, when to call a function and when to ask for more information.

Chat format and control tokens

FunctionGemma does not use a free form dialog format. It expects a robust dialog template that separates roles and regions related to tools. The turn of the conversation is wrapped with role ... where roles usually exist developer, user or model.

Within those curves, FunctionGemma depends on a fixed set of token pairs

• again for tool descriptions
• again with model tool calls
• again to the output of serialized tools

These tags allow the model to separate natural language text into functional schemas and implementation results. A Hugging Face apply_chat_template The API and official Gemma templates generate this property automatically in messages and toolbars.

Fine tuning and performance of Mobile Actions

Out of the box, FunctionGemma is already trained to use standard tools. However, the official Mobile Actions guide and model card emphasize that miniature models achieve production-level reliability only after fine-tuning a specific task.

The Mobile Actions Demo uses a dataset where each instance displays a small set of Android system functionality tools, for example create a contact, set a calendar event, control the flashlight and view the map. FunctionGemma learns to map expressions such as 'Create a calendar event for lunch tomorrow' or 'Turn on a flashlight' to those tools with set arguments.

In Mobile Action Testing, the basic FunctionGemma model achieves 58 percent accuracy on the captured test set. After fine-tuning a public cookbook recipe, the accuracy increases to 85 percent.

Edge agents and reference demos

The primary targets for FunctionGemma deployments are edge agents running locally on phones, laptops and small accelerators like the NVIDIA Jetson Nano. A small parameter count, 0.3B, and scaling support allow for low-memory and low-latency definition on consumer hardware.

Google posts several references through the Google AI Edge Gallery

• Mobile Actions demonstrates a fully offline assistant-style agent for controlling a device using FunctionGemma that has been properly configured on the Mobile Actions dataset and deployed to the device.
• The Little Garden is a voice-controlled game where the model repeats commands such as “Plant a sunflower in the upper row and water it” to perform certain background tasks such as plant_seed again water_plots with clear grid links.
• FunctionGemma Physics Playground it runs entirely in the browser using Transformers.js and allows users to solve physics puzzles by using natural language instructions that the model turns into simulated actions.

These demos confirm that the 270M parameter caller can support multi-step understanding on the device without server calls, given proper fine-tuning and tool interfaces.

Key Takeaways

1. FunctionGemma is a 270M parameter, a text-only variant of Gemma 3 that is trained to call a function, not an open dialog, and is released as an open model under Gemma's terms of use.
2. The model maintains the Gemma 3 transformer architecture and a 256k token dictionary, supports 32k tokens per request shared between input and output, and is trained on 6T tokens.
3. FunctionGemma uses a robust dialog template role ... and dedicated control tokens for function declarations, function calls and function responses, which are required for reliable tool use in production systems.
4. In the Mobile Actions benchmark, accuracy improves from 58 percent in the baseline model to 85 percent after fine-tuning a task, indicating that small task callers need domain data more than information engineering.
5. The 270M scale and scaling support allows FunctionGemma to work on phones, laptops and Jetson-class devices, and the model is already integrated into ecosystems such as Hugging Face, Vertex AI, LM Studio and edge demos such as Mobile Actions, Tiny Garden and Physics Playground.

Check it out Technical details again Model in HF. Also, feel free to follow us Twitter and don't forget to join our 100k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.