Google’s New Open LLMs: A Leap Forward in AI
Google has recently introduced two new open large language models (LLMs) named Gemma123, following closely on the heels of its latest Gemini model release
Gemma, the latest addition, comprises Gemma 2B and Gemma 7B, designed for both commercial and research purposes1. Described by Google as “state-of-the-art,” these dense decoder-only models1 draw inspiration from Gemini’s architecture
Getting Started with Gemma
Developers can dive into Gemma using pre-prepared Colab and Kaggle notebooks, with integrations available for Hugging Face, MaxText, and Nvidia’s NeMo1. Once pre-trained and adjusted, these models are versatile enough to operate across various platforms.
Open Models, Not Open Source
Despite being labeled as open models, it’s essential to clarify that they are not open-source1. Google emphasizes its commitment to open source while being deliberate in its terminology regarding the Gemma models. Developers have the freedom to utilize the models for inference and fine-tuning as needed.
Performance and Applications
Google’s team suggests that these model sizes are suitable for many applications1, with significant improvements in generation quality observed over the past year1. This advancement allows tasks previously reserved for large models to be achieved with smaller, state-of-the-art models1. This, in turn, opens up new avenues for AI application development, enabling inference and tuning on local developer machines or single hosts in GCP with Cloud TPUs.