Google releases Gemma 2

Google has taken a significant step in the evolution of its language model family with the release of Gemma 2. This extension not only introduces new models and tools, but also emphasizes the importance of security, accessibility and transparency in the development of responsible artificial intelligence (AI).

Gemma is Google's family of open source AI models, initially introduced in February. These models were designed to be lightweight and state-of-the-art, based on the same research and technology as the Gemini models.

At its original launch, the Gemma family included 2 billion-parameter (2B) and 7 billion-parameter (7B) models. In May, during its annual developer conference, Google released Gemma 2, an update with a 27 billion-parameter model (27B), offering performance on par with Meta's 70 billion-parameter Llama 3.

With the recent expansion, Google has introduced three new models within the Gemma 2 family: Gemma 2 2B, ShieldGemma and Gemma Scope. These additions are designed to prioritize security, accessibility and transparency, thus facilitating the deployment of responsible AI.

Gemma 2 new functionalities

This latest iteration of Gemma will offer three main new functionalities:

Gemma 2 2B: is a compact version of the main model, with 2 billion parameters. It incorporates significant advances in security and offers an optimal balance between performance and efficiency.

According to Google, its conversational capabilities exceed those of GPT-3.5 in Chatbot Arena.

ShieldGemma: is designed to protect users through security classifiers that identify and mitigate harmful content, such as hate speech, harassment, sexually explicit content and dangerous material, in both the inputs and outputs of AI models. This version is available in 2B, 9B and 27B sizes, the latter two being the most suitable for 'offline' applications requiring lower latency.

Gemma Scope: is a tool aimed at increasing transparency by enabling the review of Gemma 2 models' decision-making processes. It facilitates pattern identification, information processing and prediction making, improving clarity on how these AI models operate.