Google Unveils Gemma 2: A 27 Billion-Parameter Open Model Set to Launch in June

On Tuesday, at the annual Google I/O 2024 developer conference, Google unveiled exciting new additions to its Gemma family of models, which are open (but not open-source) alternatives to Meta’s Llama and Mistral’s open models. The standout reveal is Gemma 2, the next evolution of Google’s open-weight models, set to debut in June with an impressive 27 billion parameters.

In addition, Google introduced PaliGemma, a pre-trained Gemma variant that is touted as “the first vision language model in the Gemma family.” This innovative model is designed for applications such as image captioning, image labeling, and visual question-answering.

Previously, the standard Gemma models, released earlier this year, were limited to 2-billion and 7-billion parameter versions. Thus, the introduction of the 27-billion parameter model marks a significant enhancement.

During a pre-announcement briefing, Josh Woodward, Google’s VP of Google Labs, shared that the Gemma models have been downloaded “millions of times” across various platforms. He emphasized that the 27-billion model has been optimized to work seamlessly with Nvidia’s next-generation GPUs, a single Google Cloud TPU host, and the managed Vertex AI service.

However, the true value of the model lies in its performance. Although Google has yet to disclose extensive data on Gemma 2, initial impressions are promising. “We’re already seeing excellent quality; it outperforms models twice its size,” Woodward noted.

We’re launching an AI newsletter! Sign up here to start receiving it in your inboxes on June 5.