On Tuesday, during its annual Google I/O 2024 developer conference, Google unveiled several significant enhancements to its Gemma model family, which are open but not open-source, positioning them as rivals to Meta’s Llama and Mistral’s alternative models.
The most notable of these announcements is the introduction of Gemma 2, the latest iteration of Google’s open-weights Gemma models. Set to launch in June, Gemma 2 will feature an impressive 27 billion parameters, marking a substantial advancement.
Additionally, Google has made available PaliGemma, a pre-trained variant within the Gemma family. PaliGemma is described by Google as the inaugural vision-language model of the Gemma lineup, catering to applications such as image captioning, image labeling, and visual Q&A.
Until now, the Gemma models, released earlier this year, were offered in 2-billion-parameter and 7-billion-parameter configurations. The new 27-billion-parameter model represents a significant leap forward for the family.
In a pre-announcement briefing on Monday, Josh Woodward, Vice President of Google Labs, revealed that the Gemma models have been downloaded millions of times across various platforms where they are available. He emphasized that Google has optimized the 27-billion-parameter model to operate efficiently on Nvidia’s next-generation GPUs, a single Google Cloud TPU host, and the Vertex AI managed service.
However, the size of the model is not the primary concern if its performance is subpar. Google has yet to share extensive data on Gemma 2, leaving its efficacy under scrutiny until developers can fully evaluate it. “We are already witnessing impressive performance,” Woodward mentioned. “It’s outperforming models that are twice its size.”
