Google announces Gemma, a new open-source AI model

Meet Gemma, Google’s open-source AI model.

On Wednesday, Google announced a new family of models to “assist developers and researchers in building AI responsibly.” Gemma is built from the same research and technology used to create Gemini, its closed-source models that powers the Gemini chatbot (formerly Bard) and AI-related tools for Workspace (formerly Duet AI).

Will Gemma push Google ahead of the AI race?

The AI competition between Google and OpenAI is still going strong in 2024, but OpenAI appears to have the edge. Practically every other week, we’re treated with a new feature or update from Google, only for OpenAI to come out with something bigger. In December, Google announced Gemini, and less than a month later, OpenAI launched GPT stores. Last week, Google announced a “flashy” update to Gemini called “Gemini 1.5.” However, that was quickly eclipsed by OpenAI’s unveiling its AI video generator Sora.

But one thing OpenAI has yet to do is develop any open-source versions of its models (transparency isn’t really its strong suit.) Google hasn’t been entirely forthcoming about how its AI models are trained either, but it created Gemma because the company believes in “making AI helpful for everyone,” per the announcement.

Gemma comes in two model weights: Gemma 2B and Gemma 7B. Both are pre-trained and have instruction-tuned variants, which can run on a developer laptop or desktop with CPU or GPU — and Google Cloud with GPU and TPU acceleration. Unlike Gemini, which is multimodal, Gemma only has text-to-text support. But in terms of performance, Google says Gemma “surpasses significantly larger models on key benchmarks while adhering to our rigorous standards for safe and responsible outputs.” (It’s worth noting that the technical paper was not available to us at the time of publication.)

Speaking of safety and responsibility, Gemma is pre-trained to filter out personal and sensitive information. It was tested with reinforcement-learning from human feedback (RLHF) for alignment. Plus, it has undergone manual red-teaming and adversarial testing for potential harms.

Alongside Gemma, Google is also releasing a new Responsible Generative AI Toolkit, which includes safety classification, debugging, and best practice resources for developing LLMs.

Gemma is free to access on Kaggle and Colab. It’s also available through Hugging Face, MaxText, and Nvidia NeMo. Plus, first-time Google Cloud users get $300 in credits.