Google on Tuesday at its annual developer conference Google I/O 2024 announced a number of new feature additions to Gemma, a family of open (not open source) models comparable to Meta's Llama and Mistral open models.
The high-profile release here is Gemma 2, the next generation of Google's open-weight Gemma model, a 27 billion parameter model launching in June.
Already available, PaliGemma is a pre-trained Gemma variant that Google describes as “the first vision language model in the Gemma family” for image captioning, image labeling, and visual Q&A use cases.
So far, the standard Gemma model launched earlier this year only had a 2 billion parameter version and a 7 billion parameter version, but this new 27 billion model is a significant improvement.
In a press conference ahead of Tuesday's announcement, Josh Woodward, Google's vice president of Google Labs, said the Gemma model has been downloaded “millions of times” through the various services available. He highlighted that Google has optimized 27 billion models to run on Nvidia's next-generation GPUs, a single Google Cloud TPU host, and managed Vertex AI services.
However, size doesn't matter if the model is not good. Google hasn't shared much data about Gemma 2 yet, so we'll have to wait and see how it performs once developers get their hands on it. “We're already seeing great quality. It's already outperforming models twice as large,” Woodward said.
Publish an AI newsletter. Sign up here to start receiving it in your inbox on June 5th.
