At the recent Google I/O Berlin conference, Google announced the open-source release of its latest language model, Gemma2, which has made significant breakthroughs in performance and efficiency. Gemma2 offers versions with 9B and 27B parameter sizes, with the 27B version approaching the performance of the 70B parameter Llama3 model, but with only about 40% of its size.
Key features of Gemma2 include:
（1）9B and 27B parameter sizes
(2) Top-tier performance
(3) Efficient inference capabilities (runnable on a single NVIDIA H100 GPU or TPU host)
(4) An easy-to-use model designed for developers and researchers
（1）Exceptional performance: The 27B model rivals the performance of the twice-as-large Llama3 70B model
(2) High efficiency: Full-precision inference achievable on a single GPU
(3) Broad hardware support: From gaming laptops to cloud environments
(4) Open licensing: Also available for commercial use
Developer-friendly design
To facilitate developer access, Gemma2 is compatible with multiple mainstream AI frameworks such as Hugging Face, JAX, PyTorch, and TensorFlow. Google has also provided a new Gemma2 Cookbook, which includes practical application examples and guides. Additionally, Google plans to support easy deployment of Gemma2 via Google Cloud Vertex AI in the near future.
In terms of responsible AI development, Google has introduced a series of initiatives, including a responsible generative AI toolkit, an open-source LLM comparator for model evaluation, and plans to open-source the SynthID text watermarking technology. Google also commits to rigorous security assessments and the publication of their results.
Currently, developers and researchers can access the Gemma2 model for free through Google AI Studio. The model weights can also be downloaded from platforms like Kaggle and Hugging Face. For academic researchers, Google offers a Google Cloud credit program with applications due by August 9th.
It is noteworthy that Gemma2 has outperformed the QWen1.5 model on the authoritative LMSys leaderboard, further demonstrating its powerful capabilities. This groundbreaking achievement will bring new opportunities and challenges to the AI field, propelling the further development of open-source language models.
Official trial address: https://aistudio.google.com/app/prompts/new_freeform