Google launches Gemma 4, its most advanced open AI model family yet

Google has introduced Gemma 4, its most advanced open AI model family, featuring improved reasoning, multimodal capabilities and support for agentic workflows. Released under an Apache 2.0 license, it enables developers to build and deploy AI applications efficiently across devices.

Hyderabad: Google has unveiled Gemma 4, its latest generation of open artificial intelligence models, designed to deliver advanced reasoning and support agentic workflows while maintaining high efficiency across hardware platforms.

The company said Gemma 4 is its most intelligent open model family to date, offering improved performance relative to model size. Since the launch of the first-generation models, developers have downloaded Gemma more than 400 million times, creating over 100,000 variants within what Google describes as a growing “Gemmaverse”.

Built on the same research foundation as Gemini 3, Gemma 4 complements Google’s proprietary AI offerings by providing developers with flexible, open-source alternatives that can run on local hardware.

The new model family is available in four configurations: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense. According to the company, the 31B model currently ranks among the top open models globally on the Arena AI leaderboard, with the 26B model also placing within the top tier.

Google said the models are designed to move beyond basic conversational tasks, enabling complex reasoning, multi-step planning and execution of agentic workflows. Features include native support for function calling, structured outputs and system-level instructions for building autonomous AI agents.

Gemma 4 also introduces strong capabilities in code generation, allowing developers to run offline AI coding assistants on local machines. Multimodal features enable processing of images, video and, in smaller models, audio inputs for speech recognition and understanding.

The models support long context windows—up to 256K tokens in larger variants—allowing users to process extensive documents or codebases in a single prompt. Additionally, Gemma 4 is trained in more than 140 languages, supporting global application development.

Google said the models are optimised for a wide range of devices, from smartphones and laptops to high-performance computing systems. The smaller E2B and E4B variants are engineered for on-device use, offering low latency and efficient memory usage, while larger models are designed for advanced workloads on GPUs and cloud infrastructure.

The company is releasing Gemma 4 under the permissive Apache 2.0 License, allowing developers to freely build, modify and deploy applications across environments.

Google highlighted early use cases, including collaborations with INSAIT and Yale University, demonstrating applications in language modelling and biomedical research.

The models are available through platforms such as Hugging Face and Google’s own AI development tools, with support for a wide range of frameworks and deployment environments.

With Gemma 4, Google aims to strengthen its position in the rapidly evolving AI ecosystem by offering developers a combination of open and proprietary tools for building next-generation applications.