Google LLC Launches Gemma 4 Open-Weight AI Model Family

On April 2, 2026, Google LLC released Gemma 4, a family of open-weight AI models built from Gemini 3 research. The suite includes four variants tailored to different hardware: Effective 2B (E2B) and Effective 4B (E4B) for mobile and edge devices, a 26B Mixture of Experts (MoE) for high throughput, and a 31B Dense model for high-quality reasoning. The 31B model currently ranks third on the Arena AI text leaderboard.

Breaking from previous custom terms, Google LLC transitioned the line to the Apache 2.0 license to provide developers with greater flexibility and remove commercial restrictions. The models are natively multimodal, supporting text, images, video, and audio across more than 140 languages. To enable autonomous agentic workflows, Gemma 4 includes native function calling and structured JSON outputs. Smaller models feature a 128K token context window, while larger versions support up to 256K tokens.

Deployment is focused on local, offline inference to improve data privacy and reduce cloud dependence, specifically for sectors like healthcare and government. To support this, Google LLC introduced the AICore Developer Preview for Android and the AI Edge Gallery app. The launch included day-zero optimizations from Nvidia Corporation and Advanced Micro Devices, Inc. While benchmark performance is strong in mathematics, some early developer feedback indicated inconsistent inference speeds for the MoE variant and smaller context windows compared to competitors like Meta's Llama 4 Scout.