Google made good on its promise to open up its most powerful AI model, Gemini 1.5 Pro, to developers last month.
Google's Gemini 1.5 Pro can handle more complex tasks than any other AI model before it, such as analyzing entire text libraries, feature-length Hollywood movies, or an entire day's worth of audio data. That's 20 times larger than OpenAI's GPT-4o and nearly 10 times the amount of data Anthropix Cloud 3.5 Sonnet can manage.
The goal is to put faster, lower-cost tools at the disposal of AI developers, Google said in the announcement, and “enable new use cases, more product robustness and greater reliability.”
Google previously released the prototype in May, showing videos of how a group of beta testers were able to use its capabilities. For example, machine learning engineer Lucas Atkins fed the model with the entire Python library and asked questions to help solve the problem. “Nailed,” he said in the video. “He can find comments in the code and specific references to specific questions that people have asked.”
Another beta tester took video of the entire bookshelf and created a database of all the books Gemini held—a task impossible to achieve with traditional AI chatbots.
But Google is making waves in the open source community. The company today released Gemma 2 27B, an open source large language model.
Google says Gemma 2 “provides best-in-class performance, runs incredibly fast on a variety of hardware, and integrates easily with other AI tools.” It's meant to compete with models “more than twice the size,” the company says.
The Gemma 2 license allows free access and redistribution, but is still not the same as traditional open source licenses like MIT or Apache. The model is built in both 27B and and smaller 9B versions for more accessible and budget-friendly AI deployments.
This is important for average and enterprise users because, unlike the latest models, a powerful open model like Gemma is highly customizable. That means users can optimize their models for specific tasks, protecting their data by running such models locally.
For example, Microsoft's small language model Phi-3 is specifically tuned for math problems, and can beat larger models like Lama-3 and even Gemma 2 itself in that field.
Gemma 2 is now available in Google AI Studio, with model weights from Kaggle and Hug face models available for download on Vertex AI for developers with the powerful Gemini 1.5 Pro.