Google announces Gemma 4, its most powerful open source model

Gemma 4

• Google debuts Gemma 4 open models for reasoning, multimodal
• Gemma 4 runs on devices and cloud, supports 140+ languages
• Models available under Apache 2.0 license for commercial use

Did our AI summary help?

Google has announced Gemma 4, a new family of open models built for reasoning, multimodal processing, and agent-based workflows. The company said the models are designed to deliver higher performance per parameter while remaining deployable across a range of hardware, including smartphones, laptops, and cloud infrastructure. The release builds on earlier Gemma models, which have recorded over 400 million downloads.

Model sizes and performance

Gemma 4 is available in four configurations: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts, and 31B Dense. Google said the 31B model ranks among the top open models on public benchmarks, while the 26B variant focuses on lower latency by activating a smaller subset of parameters during inference.

The models are designed to balance compute requirements with performance, allowing developers to run advanced AI workloads on local systems without requiring large-scale infrastructure.

Features and capabilities

Gemma 4 introduces support for multi-step reasoning, structured outputs, and function calling, enabling the development of agent-based applications. The models also support code generation and can run locally for offline use cases.

Multimodal capabilities include support for image, video, and audio inputs, depending on the model size. The smaller E2B and E4B models are optimized for edge devices with lower memory and power requirements. Larger models offer context windows of up to 256K tokens, allowing processing of long documents and datasets.

Google also said the models support more than 140 languages, enabling broader use across global applications.

Availability and ecosystem

Gemma 4 is released under the Apache 2.0 license, allowing commercial use and modification. Developers can access the models through platforms such as Google AI Studio, Hugging Face, Kaggle, and Ollama.

The models are compatible with tools including Transformers, vLLM, MLX, and Docker, and can be deployed on local machines or scaled using Google Cloud services like Vertex AI and Kubernetes Engine.

Google said the release aims to provide developers with more flexibility to build and deploy AI systems across different environments.