Google has been riding the AI train hard these past couple of weeks. First, the company released Gemini Advanced and 1.5 Pro models, and now, the tech giant is here with Gemma AI.

Gemma is a family of lightweight, open-source AI models that are based on their flagship Gemini models. The same technologies and research have been used to build these state-of-the-art models that went into creating Gemini. You can think of Gemma as Gemini's little sibling.

What is Gemma

Designed for developers, Gemma focuses on bringing AI capabilities to everyday AI developers since they can run on various types of consumer hardware (laptops, cloud environments, or standard workstations), without the need for the extensive computational resources demanded by Gemini.

Sizes:

Gemma is currently available in two sizes: 2B and 7B (parameters), and each model has two variants, base (pre-trained) and instruction-tuned.

Google has filtered out personal information and other sensitive data from training sets to make the pre-trained models safe and reliable.

For instruction-tuned models, the company has used extensive fine-tuning and reinforcement learning from human feedback (RLHF) to make the models align with responsible behavior.

Accessing Gemma:

It's easy to get started with Gemma as it's integrated with popular tools, such as HuggingFace, Kaggle, NVIDIA NeMo, MaxText, etc. Deployment on Google Cloud is also easy through Vertex AI and Google Kubernetes Engine (GKE).

It has also been optimized for AI hardware platforms, such as NVIDIA GPUs and Google Cloud TPUs.

Availability:

Available worldwide, Gemma models will currently work in English only, with Google hoping to expand in the future. They are also best suited for language-related tasks, like question answering, summarization, and reasoning, owing to their small size.

Developers can fine tune Gemma models for their applications for tasks such as summarization or retrieval-augmented generation (RAG).

Since it's harder to put guardrails on open-sourced models, Gemma models are being shipped with responsible AI toolkits, allowing developers to create their own guidelines when using Gemma.

Gemma vs. Gemini

Gemini is available to end customers through the web app, Android app, or the Google app on iOS. But Gemma models are only designed for developers.

Developers can access Gemini through APIs or Vertex AI, making it a closed model. In comparison, Gemma is an open-source model readily available to developers, researchers, and businesses for experimentation and integration into their own applications.

Gemini models are also massive, often requiring specialized data center hardware. Whereas Gemma is smaller, making it much more portable and cost-effective to run.

While both can be fine-tuned, Gemma is built with customization in mind. Developers can more easily adapt Gemma models to work with specific types of data or perform specialized tasks.

Gemma models, especially Gemma 7B, has performed comparably in benchmark tests to other similar LLM models, like Llama 2 7B or Mistral 7B. Gemma represents Google's efforts towards making advanced AI models more accessible and adaptable. The company also plans to release more variants in the future as they expand the Gemma family.