Google has released its new AI model, the Gemma 2. As promised during the Google I/O conference, the new model was launched in June and is now available to researchers and developers. Gemma 2 is designed to provide powerful, customizable AI tools for a wide range of applications.

What is Gemma 2?

Gemma 2 is the newest generation in Google's Gemma series, created to meet the needs of developers and researchers who require powerful and manageable AI tools. This new model offers significant improvements over its predecessor, making it an excellent option for various AI applications.

Gemma 2 continues the tradition of the original Gemma series, using the same advanced research and technology found in the Gemini models. The family includes variants such as CodeGemma, RecurrentGemma, and PaliGemma, each designed for specific AI tasks. These models are easily accessible through integrations with partners like Hugging Face, NVIDIA, and Ollama.

Key Features and Capabilities

Gemma 2 is available in two sizes: 9 billion (9B) and 27 billion (27B) parameters. Both versions are designed to be lightweight and efficient, capable of running on standard gaming GPUs or a single TPU host via Vertex AI. This makes Gemma 2 a practical choice for various applications, from small projects to large-scale deployments.

Performance: The 9B model outperforms comparable models like LLaMa 3 8B, while the 27B model matches the quality of larger models such as Mixtral 8x22B.
Training Data: The 27B version is trained on 13 trillion tokens, offering robust performance, while the 9B version, distilled from a larger model, is trained on 8 trillion tokens.
Efficiency: Both models are more efficient at inference than the first generation, with significant advancements in safety features.
Accessibility: Gemma 2 can be deployed on a single NVIDIA H100 Tensor Core GPU or TPU host, significantly reducing deployment costs.

Accessing Gemma 2

To access Gemma 2, developers and researchers can use various platforms and integrations. Gemma 2 is now available in Google AI Studio (Vertex). The models are also accessible through popular AI frameworks and can be easily integrated into existing workflows. You can download Gemma 2 model weights from various sources.

For research and development, Gemma 2 is free to use on Kaggle. There is also a free tier available for Colab notebooks. First-time Google Cloud customers can get $300 in credits. Academic researchers can apply for the Gemma 2 Academic Research Program. This program offers Google Cloud credits to support research with Gemma 2. Applications are open until August 9.

Performance and Results

Initial testing of Gemma 2 on platforms like ChatBot Arena, where it was listed under the name "im-also-a-late-june-chatbot", shows promising results. Despite some variability in test data, the model demonstrates high performance and intelligence. As for now, the 9B model seem to perform better than LLaMa 3 8B, while the 27B model is as good as larger models like Mixtral 8x22B.

Chatbot Arena ELO Score:

Performance comparison versus Llama 3 and Grok1:

Conclusion

Google's Gemma 2 is set to become a valuable tool for developers and researchers seeking advanced AI capabilities. With its efficient performance, accessible deployment options, and detailed training data, Gemma 2 is expected to make a significant impact in the AI community. Stay tuned for more updates and developments as Google continues to refine and enhance the Gemma series.

While it's too early to make a final judgment, the initial outcomes are encouraging. Google is expected to address any issues that affected the original Gemma model, ensuring that Gemma 2 delivers improved performance and reliability.