Google releases Gemma open weights

Google released Gemma, a family of open-weight models — 2-billion and 7-billion parameter versions, each available pre-trained and instruction-tuned — built using the same research and infrastructure as its closed Gemini models. It was Google’s first substantial open-weight release of the generative-AI era, arriving roughly a week after the launch of Gemini 1.5.

Google said Gemma 7B outperformed Meta’s larger Llama 2 13B on standard benchmarks despite its smaller size, and released the weights under terms permitting commercial use and redistribution by organisations of any size — a licence more permissive than Llama’s, though, like Llama, not an open-source licence in the traditional sense, since it carried Google-specific usage restrictions rather than an OSI-approved licence. The release came with tooling aimed at developers: ready-made notebooks for Colab and Kaggle, and integrations with Hugging Face and NVIDIA’s inference framework, alongside a “Responsible Generative AI Toolkit” covering safety fine-tuning and content filtering.

Gemma’s release was widely read as an acknowledgement by Google that the open-weight ecosystem — driven by Meta’s Llama releases and by Mistral — had become a competitive front the company could not cede by keeping all its models closed. The models were small enough to run on a single consumer GPU or a laptop, extending distribution to hobbyists and smaller developers rather than only cloud-API customers, and set up a pattern of Google releasing successive, larger Gemma generations through 2024 and 2025 alongside its flagship closed Gemini line.