Gemma 2 2B AI Model Released by Google DeepMind, Said to Outperform GPT 3.5 Models in Benchmark

Gemma 2 2B artificial intelligence (AI) model was released by Google DeepMind on Thursday. It has become the latest addition to the Gemma 2 family of AI models and joins Gemma 2 27B and 9B models. Despite the lightweight size, the company claims that it outperformed GPT-3.5 models on the LMSYS Chatbot Arena benchmark. Alongside, the tech giant also released ShieldGemma, a suite of classifier models to filter the input and output of Gemma 2, and Gemma Scope, a research tool that offers insights into how the AI model functions.

Gemma 2 2B AI Model’s Features

In a blog post on Google for Developers, the company announced Gemma 2 2B, which has become the smallest language model of the family. Introduced as an on-device AI model, the post highlighted that despite its small parameter size, the output is significantly higher than its weight class as it was distilled from larger models. However, the tech giant did not reveal which AI models were used for its training.

Google also claimed that the Gemma 2 2B AI model outperformed the GPT-3.5 models on the large model systems organisation (LMSYS) Chatbot Arena Elo score. The AI model is said to have received a score of 1126, whereas Mixtral 8x7b Instruct v0.1 model scored 1114 and GPT-3.5 scored 1106.

The AI model has also been optimised to run on a wide range of hardware. For edge devices and cloud-based deployment, it has been fine-tuned for Vertex AI and Google Kubernetes Engine (GKE). It is also optimised for Nvidia TensorRT-LLM library and has been made available as an Nvidia NIM. Further, Gemma 2 2B also integrates with Keras, JAX, Hugging Face, and other major platforms.

Since it is an open-source AI model, the open weights can be downloaded from Google’s Hugging Face listing, Kaggle, or Vertex AI Model Garden. It can also be tried out on the Google AI Studio.

ShieldGemma
Photo Credit: Google

Apart from Gemma 2, ShieldGemma, a suite of safety classifiers that can detect and remove harmful content in both the input and output of the AI model, was also released. Google said the system will focus on hate speech, harassment, sexually explicit content, and dangerous content.

Finally, Gemma Scope, a research tool for academicians and developers was also released. The system uses sparse autoencoders (SAEs) to pinpoint specific parts within the model to highlight how the decision-making process works and how the architecture operates.