Google's latest release, the open-weight model Gemma 2, is now available to researchers and developers worldwide. The company has introduced two versions of the model, with parameter sizes of 9 billion and 27 billion, respectively.
Earlier this year, Google launched the Gemma series, a lightweight and advanced open model based on the Gemini model. Since then, this series of models has expanded, adding multiple variants such as CodeGemma, RecurrentGemma, and PaliGemma, each customized for specific AI tasks. With the recent announcement of Gemma 2 last month, Google has once again taken a significant step forward in this field.
Of particular note is the 27 billion parameter model, which performs comparably to proprietary models that are more than twice its size - something unimaginable just six months ago. Despite its significantly smaller size, it can achieve performance close to that of larger models such as Llama 3 70B and Claude 3 Sonnet.
What makes Gemma 2 particularly remarkable is its outstanding efficiency. Google claims that this model with 27 billion parameters can run in full precision on a single Google Cloud TPU host, NVIDIA A100 80GB Tensor Core GPU, or NVIDIA H100 Tensor Core GPU. This greatly reduces the hardware requirements and costs for deploying such a powerful AI model. Moreover, this high efficiency does not sacrifice speed - Gemma 2 has been optimized for fast inference on a wide range of hardware, from high-end cloud setups to consumer-grade gaming laptops.
For developers and researchers, Gemma 2 is undoubtedly a choice worth considering. It is released under a business-friendly license, allowing for research and potential monetization. Additionally, the model is designed to be highly compatible and seamlessly integrate with popular AI frameworks such as Hugging Face Transformers, JAX, PyTorch, and TensorFlow.
Currently, Gemma 2 is available in Google AI Studio and will soon be launched in the Vertex AI Model Garden. You can also download Gemma 2's model weights from Kaggle and Hugging Face Models.
