Google has unveiled Gemma 4, the latest generation of its open AI model family, marking a major step in bringing advanced artificial intelligence to both high performance data centres and everyday smartphones. The announcement reflects Google’s growing focus on making powerful AI tools more accessible, efficient and adaptable for developers worldwide.
The company revealed that since the introduction of the first Gemma models, developers have downloaded them more than 400 million times, creating a thriving ecosystem with over 100,000 variants built on top of Google’s foundation models. With Gemma 4, Google aims to push this momentum further by delivering stronger performance across a wider range of devices.
Google CEO Sundar Pichai described the new models as delivering an “incredible amount of intelligence per parameter,” highlighting the efficiency gains achieved in this release. Meanwhile, Google DeepMind CEO Demis Hassabis called Gemma 4 “the best open models in the world for their respective sizes,” emphasizing both performance and flexibility.
A unified vision across four model sizes
Gemma 4 is being released in four different configurations, each tailored to specific use cases ranging from lightweight mobile devices to powerful computing environments.
The E2B model, with effective 2 billion parameters, is designed for smartphones and IoT devices, offering efficient performance in constrained environments. The E4B model, with effective 4 billion parameters, also targets edge and mobile applications but provides additional capability for slightly more demanding tasks.
For developers needing more power, Google has introduced a 26 billion parameter Mixture of Experts model. This mid range option balances performance and latency, making it suitable for a wide variety of applications.
At the top end, the 31 billion parameter dense model stands as the flagship. According to Google, it currently ranks among the top open AI models globally and has demonstrated performance that surpasses models significantly larger in size.
This range of options reflects Google’s broader strategy to ensure that advanced AI is not limited to large scale infrastructure but can be deployed across diverse environments.
Built for performance, efficiency and flexibility
One of the defining aspects of Gemma 4 is its efficiency. Google claims the models deliver strong performance without requiring excessive computational resources. This allows developers to run advanced AI workloads even on standard hardware setups.
The models are released under the Apache 2.0 licence, enabling developers to freely use, modify and build upon them. This open approach is expected to accelerate innovation and adoption across industries.
Gemma 4 is also designed to support fine tuning, allowing developers to adapt the models for specific tasks and domains. This flexibility is particularly important for businesses looking to integrate AI into their workflows without relying entirely on proprietary systems.
Advanced capabilities beyond basic chat
Gemma 4 represents a significant upgrade in functionality compared to earlier versions. The models are capable of handling complex reasoning tasks, including multi step problem solving and improved mathematical processing.
They also support agentic workflows, enabling developers to build systems that can interact with external tools, APIs and services autonomously. This opens the door to more sophisticated AI applications that go beyond simple conversational interfaces.
Another key feature is offline code generation. Developers can run Gemma 4 locally, transforming standard machines into private AI coding assistants without requiring constant internet connectivity.
In addition, the models support multimodal inputs. All versions can process images and video, while the smaller edge models also include audio capabilities such as speech recognition.
Gemma 4 also introduces extended context windows, with smaller models supporting up to 128,000 tokens and larger models reaching up to 256,000 tokens. This allows the AI to handle longer and more complex inputs in a single interaction.
To ensure global usability, the models have been trained across more than 140 languages, making them among the most inclusive open AI systems currently available.
Optimised for smartphones and edge devices
A standout feature of Gemma 4 is its ability to run directly on consumer devices. Google has worked closely with its Pixel team as well as chipmakers like Qualcomm Technologies and MediaTek to optimise the models for mobile hardware.
The result is AI that can operate entirely offline with minimal latency, enabling faster and more private interactions. This capability extends beyond smartphones to devices such as Raspberry Pi systems and Nvidia Jetson platforms.
By bringing advanced AI directly onto devices, Google is addressing concerns around latency, data privacy and connectivity, while also unlocking new possibilities for real time applications.
Built on the foundation of Gemini technology
Google confirmed that Gemma 4 is built using the same research and technological foundations that power its flagship Gemini 3 models. This ensures that even though Gemma is open and lightweight, it benefits from cutting edge advancements in AI development.
The company’s approach reflects a dual strategy: maintaining high performance proprietary systems while simultaneously expanding access through open models like Gemma.
A significant step in democratising AI
The launch of Gemma 4 signals Google’s continued push to democratise artificial intelligence. By combining high performance with accessibility, the company is enabling developers of all sizes to build and deploy advanced AI solutions.
From enterprise scale deployments in data centres to everyday use on smartphones, Gemma 4 represents a shift toward more distributed and inclusive AI ecosystems.
As AI becomes increasingly central to digital experiences, Google’s latest release positions it strongly in the evolving landscape, where efficiency, openness and adaptability are just as important as raw performance.
