Google has introduced Gemma 2 2B, a new open-source AI model designed to excel in tasks such as text generation, translation, and question-answering. Despite having only 2 billion parameters, Gemma 2 2B is reported to outperform larger models like ChatGPT 3.5 in terms of speed and conversational capabilities. The announcement also includes the launch of new AI tools, ShieldGemma and Gemma Scope, aimed at improving safety and transparency in AI applications.

Key Features and Capabilities:

• Gemma 2 2B Model: Google’s latest AI model, Gemma 2 2B, is notable for its compact size yet high efficiency. It operates effectively across a variety of hardware, from edge devices to cloud platforms, making it versatile for numerous applications such as educational tools and personal assistants. The model is optimized for NVIDIA GPUs and is compatible with popular AI frameworks like Keras, JAX, and Hugging Face.

• ShieldGemma: This toolset includes safety classifiers designed to help developers identify and filter harmful content, including hate speech, harassment, and explicit materials. ShieldGemma promotes transparency and collaboration by being openly available, supporting Google’s commitment to safe AI practices.

• Gemma Scope: A new tool that uses sparse autoencoders to provide insights into the workings of AI models like Gemma 2. Gemma Scope simplifies complex AI decision-making processes, allowing researchers to better understand and improve model reliability and accountability.

Availability:

The Gemma 2 2B model, along with ShieldGemma and Gemma Scope, will be freely accessible. Users can download the model weights from platforms such as Kaggle, Hugging Face, and Vertex AI Model Garden. Additionally, the model can be tested on Google AI Studio and Google Colab, offering accessibility even for those with limited resources.

With these innovations, Google aims to enhance the performance and safety of AI technologies while promoting openness and collaboration within the AI community.