Google DeepMind Releases Gemma 2 2B AI Model, Which Is Said to Beat GPT 3.5 Models in Benchmark

 Gemma 2 2B is an AI model that is freely available for download from Hugging Face, Vertex AI Model Garden, and Kaggle.

Google DeepMind launched the Gemma 2 2B artificial intelligence (AI) model on Thursday. It has joined the Gemma 2 9B and 27B models as the newest member of the Gemma 2 family of AI models. The company states that on the LMSYS Chatbot Arena benchmark, it fared better than GPT-3.5 models despite its lightweight design. In addition, the tech behemoth unveiled Gemma Scope, a research tool that provides insights into the workings of the AI model, and ShieldGemma, a suite of classifier models to filter the input and output of Gemma 2.

Features of the AI Model Gemma 2 2B

 
The company unveiled Gemma 2 2B, the smallest language model in the family, in a blog post on Google for Developers. The post described an on-device AI model that was distilled from larger models, and while having a modest parameter size, produced output that was noticeably greater than its weight class. The IT behemoth did not, however, disclose which AI models were employed in its training.

Additionally, according to Google, the Gemma 2 2B AI model fared better on the large model systems organization (LMSYS) Chatbot Arena Elo score than the GPT-3.5 models. The AI model reportedly scored 1126, while the GPT-3.5 and Mixtral 8x7b Instruct v0.1 models scored 1114 and 1106, respectively.

Additionally, the AI model has been tuned to function on a variety of hardware. It is optimized for Vertex AI and Google Kubernetes Engine (GKE) for edge devices and cloud-based deployment. It has also been released as an Nvidia NIM and optimized for the Nvidia TensorRT-LLM library. Additionally, Gemma 2 2B connects with Hugging Face, JAX, Keras, and other significant platforms.

Since the AI model is open-source, you can get the open weights via Vertex AI Model Garden, Kaggle, or Google's Hugging Face listing. You can test it out on the Google AI Studio as well.

In addition to Gemma 2, ShieldGemma is a set of safety classifiers that was also released. It has the ability to identify and eliminate dangerous content from both the input and output of the AI model. According to Google, the system will prioritize information that is harmful, sexually explicit, hateful, and harassing.

Lastly, a research tool for developers and academics called Gemma Scope was also made available. Sparse autoencoders (SAEs) are used by the system to identify certain areas of the model that demonstrate the functioning of the architecture and the decision-making process.