On February 21, Google announced Gemma 2B and 7B, open-source artificial intelligence models based on Gemini. Gemini line – neural networks used for various purposes: Gemini Nano, Gemini Pro and Gemini Ultra. Gemini 1.5 was also recently announced, being the faster version, but so far only for enterprises and developers.
Thus, Gemini Ultra can recognize, analyze and generate texts, images, audio and video. With Gemini Pro, developers can work in their preferred environment: SDKs are available for Python, Android (Kotlin), Node.js, Swift and JavaScript.
Unlike Gemini, accessible through the API or Vertex AI, Gemma aims to attract a wider range of developers.
Let’s move on to the key points of the presentation:
- Gemma is a family of models. There will be two models in sizes 2B and 7B.
- Gemma will use the new Responsible Generative AI Toolkit to help prioritize the creation of AI apps.
- The multi-framework Keras 3.0 provides compatibility with JAX, PyTorch and TensorFlow, which will allow developers to quickly switch platforms depending on their tasks.
- Gemma is equipped with popular tools such as Hugging Face, MaxText, NVIDIA NeMo and TensorRT-LLM, and also uses Colab and Kaggle notebooks.
- Gemma’s pre-trained and tuned models can run on your laptop, PC, or Google Cloud with easy deployment on Vertex AI and Google Kubernetes Engine (GKE).
- Gemma boasts improved performance in a modest size, including NVIDIA GPUs and Google Cloud TPUs. Vertex AI provides a variety of MLOps tools with one-click configuration and deployment using built-in output optimizations.
- Google says its Terms of Service will allow all organizations, regardless of size, to use Gemma. However, the tool is currently only suitable for English-speaking use.
- The models will be free to use on the Kaggle platform, and new Google Cloud customers will be able to get a $300 discount on their deployment. For researchers, its size reaches $500 thousand.
Google emphasizes that Gemma becomes the best due to its size, the ability to run directly on a laptop or PC, and high key indicators. You can read more details about its performance in Google’s technical report.
Developers actively use neural networks when writing code, and in addition to Google tools, we found several more popular neural assistants: Copylot on the OpenAI Codex model, the well-known ChatGPT, Fig – a very useful tool for beginners who have not yet mastered all the functionality of programming languages and development patterns, Mintlify – that helps write code documentation.
The emergence of new technologies and neural tools always contributes to the evolution of development, and Gemma is no exception, offering developers new perspectives and opportunities.