Google DeepMind has released the Gemma 2 2B AI model, which is reported to outperform GPT-3.5 models in benchmark tests
The Gemma 2 2B is an open-source AI model available for download from Kaggle, Hugging Face, and the Vertex AI Model Garden.
On Thursday, Google DeepMind released the Gemma 2 2B AI model, the newest member of the Gemma 2 family, which includes the Gemma 2 27B and 9B models. Despite its smaller size, the company claims that it outperforms GPT-3.5 models on the LMSYS Chatbot Arena benchmark. In addition, Google DeepMind introduced ShieldGemma, a suite of classifier models designed to filter the inputs and outputs of Gemma 2, and Gemma Scope, a research tool that provides insights into the functioning of the AI model.
Gemma 2 2B AI Model’s Features
In a blog post on Google for Developers, the company introduced the Gemma 2 2B, the smallest language model in the Gemma 2 family. Marketed as an on-device AI model, the post noted that despite its compact size, its performance significantly exceeds expectations due to being distilled from larger models. However, the specific models used for training were not disclosed.
Google also reported that the Gemma 2 2B AI model surpassed GPT-3.5 models in the LMSYS Chatbot Arena Elo score. The Gemma 2 2B achieved a score of 1126, while the Mixtral 8x7b Instruct v0.1 model scored 1114 and GPT-3.5 scored 1106.
The AI model has been optimized to run across a variety of hardware platforms. It is fine-tuned for use with Vertex AI and Google Kubernetes Engine (GKE) for edge devices and cloud-based deployments. Additionally, it is optimized for the Nvidia TensorRT-LLM library and is available as an Nvidia NIM. Gemma 2 2B also integrates with major platforms such as Keras, JAX, and Hugging Face.
As an open-source AI model, its weights can be downloaded from Google’s listings on Hugging Face, Kaggle, or the Vertex AI Model Garden. It is also available for experimentation on the Google AI Studio.
In addition to Gemma 2, Google also released ShieldGemma, a suite of safety classifiers designed to detect and remove harmful content from both the input and output of the AI model. The system is aimed at addressing issues related to hate speech, harassment, sexually explicit content, and dangerous material.
Finally, Google released Gemma Scope, a research tool for academics and developers. This system utilizes sparse autoencoders (SAEs) to identify specific components within the model, providing insights into the decision-making process and the model’s architecture.
