Google open sourced its latest language model Gemma2 at the Berlin I/O conference, which has attracted widespread attention for its excellent performance and efficiency. Gemma2 is available in 9B and 27B parameter sizes, with even the smaller 9B version showing impressive performance. It is particularly worth noting that the performance of the 27B version is close to the Llama3 model with 70B parameters, but it is only 40% of its size, which is a significant breakthrough in model efficiency. Gemma2's open source and compatibility with multiple AI frameworks will provide developers and researchers with a powerful tool to promote further development in the field of AI.
At the recent Google I/O Berlin conference, Google announced that it would open source its latest language model Gemma2, which has achieved major breakthroughs in performance and efficiency. Gemma2 provides two versions with 9B and 27B parameter sizes. The performance of the 27B version is close to the Llama3 model with 70B parameters, but the model size is only about 40% of the latter.
Key features of Gemma2 include:
(1) Provide 9B and 27B parameter sizes
(2) First-class performance
(3) Capable of efficient inference (running on a single NVIDIA H100GPU or TPU host)
(4) Easy-to-use models designed for developers and researchers
In addition, Gemma2 has the following features:
(1) Excellent performance: The 27B model is comparable to the Llama3 70B model which is more than 2 times the size
(2) High efficiency: a single GPU can achieve full-precision inference
(3) Broad hardware support: from gaming laptops to cloud
(4) Open license: also available for commercial use
Developer-friendly design
For the convenience of developers, Gemma2 is compatible with a variety of mainstream AI frameworks, such as Hugging Face, JAX, PyTorch, and TensorFlow. Google also provides a new Gemma2Cookbook with practical application examples and guidance. 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 launched a series of initiatives, including providing a responsible generative AI toolkit, open source LLM comparator (for model evaluation), and plans to open source SynthID text watermarking technology. Google also pledged to conduct rigorous security assessments and publish the results.
Currently, developers and researchers can obtain the Gemma2 model for free through Google AI Studio. Model weights can also be downloaded from Kaggle and Hugging Face platforms. For academic researchers, Google also offers a Google Cloud Credits program, with an application deadline of August 9.
It is worth noting that Gemma2 outperformed the QWen1.5 model on the authoritative LMSys list, further proving its powerful performance. This breakthrough achievement will bring new opportunities and challenges to the AI field and promote the further development of open source language models.
Official experience address: https://aistudio.google.com/app/prompts/new_freeform
All in all, the open source of Gemma2 marks a new stage in the development of open source large-scale language models. Its efficient performance and ease of use will accelerate the popularization and application of AI technology and bring more possibilities to the AI community. We look forward to Gemma2 being able to play a role in more fields in the future.