When Google DeepMind launched the Gemma model family in February 2024, the goal was to offer an open-weight alternative to proprietary AI models. Sixteen months later, Gemma crossed 900 million total downloads. Gemma hit 150 million downloads by May 2025, and by April 2026, that number had reached 500 million. Gemma 4 launched in April 2026, and its variants alone have accounted for over 300 million downloads since then. The family includes models ranging from 2 billion to 27 billion parameters, with specialized variants like ShieldGemma and MedGemma. Hugging Face hosts more than 70,000 fine-tuned variants of Gemma models. Google’s motivation includes using the model family as an on-ramp to capture cloud infrastructure spend. The 70,000-plus fine-tuned variants represent distributed R&D, but the open weights mean Google cannot control how the models are used, and some variants will directly compete with Google’s own commercial offerings.