Google Releases Gemma 4 Open Models Under Apache 2.0 License

Apr 3, 2026

Google's DeepMind division released Gemma 4 on April 2, 2026, marking the first time Google has used the fully open Apache 2.0 license for its AI models instead of its previous restricted Gemma Terms of Use. The release includes four model variants ranging from 2-billion parameter edge models (E2B, E4B) designed for devices like Raspberry Pi and phones, to larger 26B and 31B models for workstations and servers, all supporting multimodal capabilities across video, images, and audio. The models support up to 256K context windows, native training on over 140 languages, and enable completely offline deployment that keeps data on-premises—critical for healthcare providers and enterprises with regulatory compliance requirements. Since the first Gemma generation launched in February 2024, developers have downloaded Gemma models over 400 million times and created more than 100,000 variants, demonstrating significant adoption in the open AI ecosystem.

Gemma 4's Apache 2.0 Open-Source Release

▪Google announced on April 2, 2026 that its DeepMind AI research division is releasing Gemma 4 under the Apache 2.0 license.
▪The Apache 2.0 license allows users and developers to use Gemma 4 software for any purpose without royalty requirements.

Local AI Deployment Capabilities and Use Cases

▪Gemma 4 enables offline, multimodal AI deployment on servers, phones, and Raspberry Pi devices.

Technical Specifications and Model Variants

▪The Gemma 4 31B model is designed to maximize raw power and quality by bringing all its capabilities to any problem.
▪The Gemma 4 E2B and E4B models feature native audio input for speech recognition and understanding.
▪The Gemma 4 E2B and E4B models have two-billion and four-billion parameter footprints respectively and are designed for mobile and IoT devices.
▪The Gemma 4 E2B and E4B models support a 128K context window.
▪Gemma 4 E2B and E4B models run completely offline with near-zero latency on edge devices like phones, Raspberry Pi, and Jetson Nano.
▪All Gemma 4 models natively process video and images, support variable resolutions, and excel at visual tasks like OCR and chart understanding.
▪The Gemma 4 26B model focuses on reducing latency by activating a subset of its total parameter set for inference.

Gemma Ecosystem Growth and Adoption

▪The Gemma ecosystem includes more than 100,000 variants built by developers.
▪Google released the first generation of Gemma in February 2024.
▪Developers have downloaded Gemma over 400 million times since the first generation launch.