Google Just Open-Sourced Its Agent Models Under Apache 2.0, and Your Multi-Agent System Leaks Secrets by Design

Google released Gemma 4, its most capable open models, under a full Apache 2.0 license. The family includes a 31B dense model and a 26B Mixture-of-Experts model, both with 256K-token context windows, plus E2B and E4B edge models that run on phones and laptops with audio, image, and text support. The Gemma family has been downloaded over 400 million times, with more than 100,000 community variants.

The license shift matters more than the benchmarks. Apache 2.0 means enterprises can deploy, modify, and commercialize without the restrictions that came with earlier Gemma releases. When the company with the largest AI infrastructure makes its agentic models fully open-source, it resets the cost floor for every startup building agent products. Google is betting that commoditizing the model layer accelerates adoption of its cloud and tooling ecosystem. For builders, the implication is practical: production-grade agentic models now run locally on consumer hardware, with no API dependency and no usage fees. (Google AI Blog · VentureBeat · Google DeepMind)