Google Releases Gemma 4 Family of Five Open Models for Mobile to Workstation Use

Google DeepMind launched the Gemma 4 family of open models under the Apache 2.0 license, releasing four sizes on March 31, 2026, and adding a fifth — the 12B Unified — on June 3, 2026. The lineup spans five sizes from the 2B edge model to the 31B dense model, all supporting text and image input with context windows of 128K or 256K tokens across more than 140 languages. A standout model is the 26B A4B, a Mixture-of-Experts architecture that activates only around 4B parameters per token, delivering near-31B quality at a fraction of the memory cost at approximately 14.4GB quantized. The 12B Unified model eliminates separate encoders entirely, processing text, images, and audio through a single architecture at roughly 6.7GB in 4-bit quantization. Google's own benchmarks show significant gains over the previous Gemma 3 27B, with the 31B model scoring 89.2 on AIME 2026 compared to 20.8 for its predecessor, though these figures reflect Google's own testing conditions.