Released under the business-friendly Apache 2.0 licence, the new family of models can be used freely in both academic and commercial products, including fully local deployments.

Available in four variants, Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts and 31B Dense, Gemma 4 scales from smartphones and edge devices to powerful workstations. The larger models support context windows of up to 256,000 tokens, while the smaller edge-focused versions offer 128,000-token context, enabling long documents and complex workflows to be handled in a single run. Google says the system has been trained natively on data covering more than 140 languages, with strong multilingual performance out of the box.

Google describes Gemma 4 as its most capable open model line so far, with a clear emphasis on reasoning and "agentic" behaviour. It supports multi-step planning and deep logic, with noticeable gains on mathematics and instruction-following benchmarks, and includes native function-calling and structured JSON output to help it drive autonomous software agents. Those agents can call tools and APIs, coordinate workflows and act on behalf of users, rather than simply respond to chat prompts.

The models also support high-quality offline code generation, turning a laptop or desktop into a local coding assistant, and integrate with platforms such as Google AI Studio and Vertex AI, as well as repositories including Hugging Face, Kaggle and Ollama.

All Gemma 4 models can process images and video natively, handling tasks such as optical character recognition and chart understanding at variable resolutions. The E2B and E4B variants go further by adding audio input for speech recognition and speech-to-translated-text across multiple languages, making them suitable for on-device assistants and transcription tools.

By combining open weights, a permissive licence, long-context reasoning and multimodal input, Google Gemma 4 open-source AI model is designed to give developers far more control over data, infrastructure and deployment than typical cloud-only services, while still pushing local AI closer to state-of-the-art performance.