Google Launches Open-Source Model Google Gemma 2 for High-Speed Inference on Standard Computers
Brook.X
Google's AI team has recently unveiled Google Gemma 2, a significant update in the Google Gemma series of open-source models. This version not only boasts a smaller size to fit standard hardware but also provides exceptional capabilities.
The open-source and open model landscape features numerous models including Meta's Llama series, Alibaba Cloud's Qwen series, and Google's Gemma series. However, it's crucial to distinguish that Meta's Llama is open but not open-source, whereas both Alibaba's Qwen and Google's Gemma are truly open-source. For instance, Gemma 2 is available under the Apache License v2.0, allowing individuals, developers, and businesses to freely use these models to advance AI technology.
Focusing on the Gemma 2 model, Google offers 9B and 27B versions that can run on high-performance laptops, desktops, and even cloud-based systems. They are designed to be operational with minimal hardware requirements, such as a single NVIDIA H100 AI accelerator card, significantly lowering the barrier for running high-performance models.
Key Advantages of the Gemma 2 Model:
- Superior Capabilities: The 27B version of Gemma 2 surpasses other models of similar size in capability, rivaling even those with twice as many parameters.
- Efficiency and Cost Savings: The 27B version can be run on Google Cloud TPU hosts, NVIDIA A100 80GB, and NVIDIA H100, enhancing inference capabilities.
- Ultra-Fast Inference Across Hardware: Thanks to its redesigned architecture, Gemma 2 achieves high-speed performance across various hardware, including laptops and desktops.
Google is considering integrating the Gemma model into the Chrome browser, potentially upgrading Chrome to use the 9B or 27B versions of Gemma 2 to empower developers. These models can run entirely offline, allowing developers to build applications that leverage AI capabilities.
With an open-source license, developers and businesses can download, fine-tune, and use Gemma 2 to build their applications. Starting next month, Google Cloud computing users will easily deploy and manage Gemma 2 on Vertex AI. However, developers and businesses can already test Gemma 2 on Google AI Studio and download the model from Hugging Face: http://huggingface.co/google/
Google offers various versions of the model on Hugging Face, including Gemma 2 9B, Gemma 2 9B IT, Gemma 2 27B, and Gemma 2 27B IT, allowing developers to choose according to their needs.
