Google has introduced its latest AI model, Gemma 2 2B, which is supposed to be lightweight while yet giving great performance. According to reports, the new model outperforms larger models such as GPT-3.5 and Mistral 8x7B on a variety of metrics. This release comes after the debut of Gemma 2. The Gemma 2 2B model adds more safety features and is part of a larger update that includes ShieldGemma and Gemma Scope.
Gemma 2 2B is designed to be a lightweight model that uses knowledge from larger models to get substantial outcomes. As per Google, it outperforms all GPT-3.5 models in Chatbot Arena, displaying significant conversational AI capabilities.
Google in its official release stated, “With these additions, researchers and developers can now create safer customer experiences, gain unprecedented insights into our models, and confidently deploy powerful AI responsibly, right on the device, unlocking new possibilities for innovation.”
The approach Google has opted for, is intended to work with a wide range of hardware, from edge devices and laptops to large-scale cloud systems that employ Vertex AI and Google Kubernetes Engine (KGE). It is optimised for the NVIDIA TensorRT-LLM library and works with tools and platforms such as Keras, JAX, Hugging Face, NVIDIA NeMo and Ollama. It will also be available via the MediaPipe platform.
The Gemma 2 2B model includes 2.6 billion parameters and was trained using a dataset of 2 trillion tokens. It scored 1130 on Chatbot Arena, which is comparable to GPT-3.5 Turbo and Mixtral - 8x7b. The model also scored 56.1 in Massive Multitask Language Understanding (MMLU) and 36.6 in Mostly Basic Python Problems (MBPP), representing a performance boost of more than 10 per cent over its predecessor.
Gemma 2 2B, an open-source model, is available for download from Google's announcement page, making it easy to integrate into a variety of applications. This release represents Google's continued efforts to expand AI technology and make solutions accessible to developers and researchers.