A specialized AI model that’s been newly added to the Google Gemma family is optimized for developers that need maximum efficiency to execute high-volume, well-defined tasks such as sentiment analysis, content moderation, or compliance checking.
Gemma 3 270M has a total of 270 million parameters: 170 million embedding parameters and 100 million parameters for transformer blocks. With a vocabulary of 256,000 tokens (roughly equivalent to 192,000 words), the model can handle various types of tokens, making it a strong base model to be fine-tuned for specific domains and languages.
Gemma 3 270M follows recent additions to the Gemma lineup that include Gemma 3 and Gemma 3 QAT, which provide optimal performance for single cloud and desktop accelerators, respectively. Also added was Gemma 3n, a mobile-first model providing real-time multimodal AI directly on edge devices. Gemma downloads recently surpassed 200 million.
Right Tool, Right Job
The latest Google model features strong capacity to follow verifiable instructions for on-device and research applications. It’s not designed for complex conversational use cases; instead, its strength is following general instructions out of the box.
In addition to its compact architecture, Gemma 3 270M addresses a vital consideration for enterprises and developers with high energy efficiency and low power consumption. Google offered a highly specific test result for its use on the company’s Pixel 9 Pro phone: Gemma 3 270M used a mere .75% of the battery capacity while handling 25 conversations. The company said this is the most power efficient model in the Gemma family.
It also has quantization checkpoints available, letting developers run the models with precision and minimal performance degradation – that’s an important consideration for its use on resource-constrained devices. Quantization checkpoints are designed to reduce the model’s memory footprint and increase inference speed by converting weights and activations to lower numerical precision.
Google outlined several scenarios where Gemma 3 270M could be an optimal solution for developers with:
- Lightweight, inexpensive infrastructure or operating directly on a device. Gemma 3 270M will reduce inference costs while delivering fast responses.
- A need to iterate and deploy quickly. The model support rapid fine-tuning experiments to quickly identify an optima configuration for your use case.
- Paramount user privacy considerations; since the model can run entirely on a device, developers can build applications handling sensitive information without sending data to the cloud
- A need for multiple specialized task models; they can build and deploy multiple custom models trained for specific tasks.
More AI in Search
Google also added agentic capabilities in the AI Mode of Google Search. With the new features, AI Mode can be used to find restaurant reservations and will soon expand to support local service appointments and event tickets.
Regarding the dinner reservation capability, consumers can input multiple constraints and preferences (date, time, location, 0cuisine) and AI Mode will search across reservation platforms, finding real-time availability for those restaurants that meet the specific requirements.
AI Mode uses the live web browsing capabilities of Project Mariner, direct partner integrations (OpenTable, Resy, Tock, Ticketmaster, StubHub, SeatGeek), as well as Google Knowledge Graph and Google Maps to help users take action.
This new agentic experience is rolling out for Google AI Ultra subscribers in the U.S. through the Labs experiment named Agentic capabilities in AI Mode.
AI Agent & Copilot Summit is an AI-first event to define opportunities, impact, and outcomes with Microsoft Copilot and agents. Building on its 2025 success, the 2026 event takes place March 17-19 in San Diego. Get more details.
