NVIDIA and Google DeepMind Unveil Optimized Gemma 3 AI Models
The collaboration between NVIDIA and Google DeepMind has resulted in the introduction of the Gemma 3 series, a novel range of AI models designed to cater to a variety of computing environments. These models are engineered for optimal performance, offering developers flexible options for deploying generative AI capabilities across diverse platforms, from data centers to edge computing systems, according to NVIDIA.
Versatile AI Models for Diverse Applications
The Gemma 3 lineup includes a 1B text-only small language model (SLM) and three image-text models of sizes 4B, 12B, and 27B. These models are available on platforms such as HuggingFace and NVIDIA's API Catalog, allowing developers to experiment with high-quality, customizable models that support large-scale services across various environments.
The 1B model is particularly optimized for devices requiring low memory usage, supporting inputs of up to 32K tokens, while the larger models can handle text and image inputs up to 128K tokens. This adaptability makes Gemma 3 models suitable for both on-device applications and high-demand tasks.
Enhanced Experimentation and Prototyping
Developers can explore and prototype with Gemma 3 models using the NVIDIA API Catalog, where they can adjust parameters like max tokens and sampling values. The platform generates code in Python, NodeJS, and Bash, facilitating integration into existing workflows. Additionally, NVIDIA's LangChain library supports the creation of reusable clients for building AI agents and connecting external data sources.
Support for Robotics and Edge Solutions
The Gemma 3 models are compatible with NVIDIA's Jetson family of embedded computing boards, which are widely used in robotics and edge AI applications. The smaller 1B and 4B models can operate on compact devices like the Jetson Nano, while the more robust 27B model is optimized for the Jetson AGX Orin, offering up to 275 TOPS for high-demand applications.
Ongoing NVIDIA and Google Collaboration
Both companies have collaborated extensively on the development of Gemma models, with NVIDIA optimizing them for GPUs and contributing to the advancement of machine learning libraries like JAX and compilers such as Google's XLA. This partnership underscores a commitment to enhancing AI transparency and resilience.
NVIDIA’s contributions to open-source projects further support these efforts, with the Gemma models being part of a broader initiative to encourage AI safety and customization across industries through platforms like NVIDIA NeMo.
For developers interested in exploring these new models, Gemma 3 can be accessed and tested through the NVIDIA API Catalog, providing a robust platform for the next generation of AI solutions.