Google has unveiled the latest iteration of its lightweight, state-of-the-art open model series, Google Gemma 2.

Designed to run efficiently on NVIDIA’s latest GPUs or a single TPU host in Vertex AI, Gemma 2 promises to deliver exceptional performance and cost efficiency for developers and researchers globally.

Model Availability

Google Gemma 2 is available in two sizes

Performance and Efficiency

Gemma 2 is engineered to provide superior performance for its size, delivering class-leading results in AI tasks.

Notably, the 27B model offers performance competitive with models more than twice its size, making it a powerful alternative to proprietary models.

Training and Evaluation

Google has ensured the robustness and safety of Gemma 2 through meticulous processes:

• Filtering Pre-Training Data: To eliminate biases.

• Rigorous Testing and Evaluation: To identify and mitigate potential risks.

• Comprehensive Metrics: Used to assess and improve model safety and performance.

Key Features of Gemma 2

Outsized Performance

27B Model: Delivers top performance for its size class, rivaling much larger models.

9B Model: Outperforms competitors like Llama 3 8B, setting a new standard in its category.

Unmatched Efficiency and Cost Savings

Single GPU or TPU Host Deployment: The 27B model runs efficiently on a single Google Cloud TPU host, NVIDIA A100 80GB Tensor Core GPU, or NVIDIA H100 Tensor Core GPU, significantly cutting costs.

Blazing Fast Inference

Hardware Optimization: Gemma 2 is optimized for rapid inference across various hardware, from gaming laptops and desktops to cloud-based setups.

Full Precision Inference: Available in Google AI Studio, and a quantized version with Gemma.cpp for local performance.

Developer and Researcher Friendly

Open and Accessible

Commercially-Friendly License: Allows developers and researchers to share and commercialize their innovations freely.

Broad Framework Compatibility

Integration with Major AI Frameworks: Compatible with Hugging Face Transformers, JAX, PyTorch, TensorFlow, Keras 3.0, and more.

NVIDIA Optimization: Includes support for NVIDIA TensorRT-LLM and NeMo, enabling efficient deployment on NVIDIA-accelerated infrastructure.

Effortless Deployment

Vertex AI Integration: Starting next month, Google Cloud customers can deploy and manage Gemma 2 seamlessly on Vertex AI.

Responsible AI Development

Google emphasizes the importance of responsible AI development, providing tools and resources for safe deployment:

Responsible Generative AI Toolkit: Includes the open-sourced LLM Comparator for in-depth model evaluation.

Text Watermarking Technology: SynthID will be open-sourced for use with Gemma models.

Success Stories and Future Prospects

The initial Gemma launch saw over 10 million downloads, with innovative projects like Navarasa, which focused on India’s linguistic diversity. With Gemma 2, developers can achieve even more ambitious goals, leveraging new levels of performance and potential.

Getting Started with Gemma 2

Gemma 2 is now accessible through:

• Google AI Studio: Test its full performance capabilities at 27B without hardware requirements.

• Kaggle and Hugging Face Models: Download Gemma 2’s model weights.

• Vertex AI Model Garden: Coming soon.

For researchers, Gemma 2 is available free of charge through Kaggle or a free tier for Colab notebooks. Additionally, academic researchers can apply for the Gemma 2 Academic Research Program to receive Google Cloud credits.

Conclusion

Google Gemma 2 is set to redefine efficiency and performance in AI development, offering powerful tools for developers and researchers. With its robust design, broad compatibility, and commitment to responsible AI, Gemma 2 is poised to unlock new possibilities in AI innovation.

Start exploring the capabilities of Gemma 2 today and join the revolution in AI development.

If you want more updates related to AI, subscribe to our Newsletter