Gemma 4 Release Notes: High-Intelligence Open Models for Reasoning and Agents

Google DeepMind has introduced Gemma 4, a family of open models purpose-built for advanced reasoning and agentic workflows. These models are designed to provide high intelligence-per-parameter, allowing developers to achieve frontier-level capabilities on a wide range of hardware, from mobile devices to professional workstations.

Model Variants and Performance

Gemma 4 is released in four distinct sizes to balance raw intelligence with hardware efficiency:

• 31B Dense: Designed for maximum quality and as a foundation for fine-tuning. It currently ranks as the #3 open model globally on the Arena AI text leaderboard.
• 26B Mixture of Experts (MoE): Optimized for low latency, activating only 3.8 billion parameters during inference. It ranks #6 on the Arena AI text leaderboard.
• Effective 4B (E4B): An edge-optimized model prioritizing multimodal capabilities and low-latency processing for mobile and IoT devices.
• Effective 2B (E2B): A highly efficient edge model designed for minimal RAM and battery consumption on devices such as phones and Raspberry Pi.

Key Technical Capabilities

Gemma 4 introduces several architectural and functional improvements to support complex AI applications:

Advanced Reasoning and Agentic Workflows

The models are capable of multi-step planning and deep logic, showing significant gains in math and instruction-following benchmarks. To support autonomous agents, Gemma 4 includes native support for function-calling, structured JSON output, and native system instructions.

Multimodal Processing

All Gemma 4 models natively process video and images, supporting variable resolutions and tasks such as OCR and chart understanding. The E2B and E4B edge models further include native audio input for speech recognition and understanding.

Context Window and Language Support

The larger 26B and 31B models offer a context window of up to 256K tokens, while the edge models (E2B and E4B) feature a 128K context window. The family is natively trained on over 140 languages.

Code Generation

Gemma 4 supports high-quality offline code generation, enabling local-first AI coding assistants.

Hardware Optimization and Deployment

Personal Computers and Workstations

The unquantized bfloat16 weights for the 26B and 31B models fit on a single 80GB NVIDIA H100 GPU. Quantized versions are available for consumer GPUs to support local IDEs and agentic workflows.

Mobile and IoT Edge Devices

The E2B and E4B models are engineered for maximum memory efficiency to run offline with near-zero latency on Android devices, Raspberry Pi, and NVIDIA Jetson Orin Nano. Android developers can prototype these flows via the AICore Developer Preview for compatibility with Gemini Nano 4.

Licensing and Ecosystem

Apache 2.0 License

Gemma 4 is released under a commercially permissive Apache 2.0 license, providing developers with complete control over their data, infrastructure, and models.

Integration and Tooling

Gemma 4 is supported on day one across a wide array of platforms and tools, including:

• Model Hubs: Hugging Face, Kaggle, and Ollama.
• Frameworks: Transformers, TRL, vLLM, llama.cpp, MLX, and Keras.
• Deployment: Google AI Studio, Vertex AI, Cloud Run, GKE, and NVIDIA NIM.
• Hardware: NVIDIA (Jetson Orin Nano to Blackwell GPUs), AMD (via ROCm stack), and Google TPUs (Trillium and Ironwood).

"The release of Gemma 4 under an Apache 2.0 license is a huge milestone. We are incredibly excited to support the Gemma 4 family on Hugging Face on day one."

— Clément Delangue, co-founder and CEO, Hugging Face