Google launches Gemma 2, its open-weight model family, in 9B and 27B sizes

Google released Gemma 2, the second generation of its open-weight model family built on Gemini research, in 9-billion and 27-billion parameter sizes (a smaller 2.6B variant was announced as forthcoming). The company said the models were more efficient at inference and had received expanded safety work compared with the original Gemma line launched in February 2024.

Google’s own comparisons put the 9B model ahead of Meta’s Llama 3 8B among similarly sized open models, and described the 27B model as offering “competitive alternatives to models more than twice its size” — a claim aimed at closed and larger open competitors alike. The 27B model was designed to run on a single Nvidia H100 GPU or a TPU host, which Google said meaningfully lowered the cost of deployment compared with models of similar capability.

Weights were released under Google’s Gemma licence, which permits commercial use and redistribution of derivative work, and the models were made available through Google AI Studio, Kaggle and Hugging Face at launch, with Vertex AI support following the next month. Gemma 2 continued a pattern through 2024 in which frontier labs released smaller open models trained on techniques distilled from their flagship systems, competing on efficiency and cost-per-token rather than on outright scale.