What Happened
Google released Gemma 4 on April 2, 2026, with downloadable weights under Apache 2.0. The initial family contained Effective 2B, Effective 4B, a 26B mixture-of-experts model and a 31B dense model. Google’s announcement identified immediate availability and linked to the released models.
Why It Matters
Gemma 4 combined a major capability update with permissive licensing and deployment options spanning phones, laptops and servers. That combination widened opportunities to adapt multimodal reasoning models and run them on independently controlled infrastructure. Native tools and structured outputs also made the family relevant to local agent applications.
Technical Details
All four initial variants process images and video; E2B and E4B also accept audio. The smaller variants provide 128K context, while the larger models support up to 256K. The 26B mixture-of-experts variant activates 3.8 billion parameters per inference step. The family supports function calling, JSON output and configurable reasoning. E2B and E4B describe effective parameter footprints.