Google has made a significant announcement in the field of artificial intelligence (AI) with the introduction of Gemma, a new family of AI language models. These models, which are free and open, are built on technology similar to Google’s more powerful but closed Gemini models. The key difference is that Gemma models can run locally on a desktop or laptop computer, providing users with greater accessibility and control.
Gemma models come in two sizes: Gemma 2B with 2 billion parameters and Gemma 7B with 7 billion parameters. Both sizes are available in pre-trained and instruction-tuned variants. Parameters are values in a neural network that determine the behavior of an AI model, while weights are a subset of these parameters stored in a file. Developed by Google DeepMind and other Google AI teams, Gemma draws from the techniques learned during the development of Gemini, Google’s most capable commercial large language models (LLMs).
This release marks Google’s first major open LLM since the launch of OpenAI’s ChatGPT, which sparked a frenzy for AI chatbots in 2022. While Google has previously released smaller research models like FLAN-T5, Gemma represents a significant contribution to the field of open AI research. Google has been actively involved in advancing AI technology, with notable contributions such as the development of the Transformer architecture and releases like TensorFlow, BERT, T5, and JAX.
Smaller open-weights LLMs have often been seen as tech demos due to their limited capability and high confabulation rates. However, larger models have started to match the performance levels of GPT-3.5. Experts believe that source-available and open-weights AI models are crucial for ensuring transparency and privacy in chatbots. While Gemma is not technically “open source,” it still represents a step towards greater accessibility and control for developers.
The introduction of Gemma can also be seen as a response to Meta’s efforts in releasing open-weights models like LLaMA and Llama 2. Meta has emphasized the importance of open AI models, while Google’s Gemini models have been limited to its own AI assistant. Gemma’s release aims to level the playing field and attract more developers to Google’s Vertex AI cloud platform.
Google claims that the Gemma 7B model outperforms Meta’s Llama 2 7B and 13B models on various benchmarks, including math, Python code generation, general knowledge, and commonsense reasoning tasks. The Gemma models are available today through Kaggle, a machine-learning community platform, and Hugging Face.
In addition to the Gemma release, Google has also unveiled a “Responsible Generative AI Toolkit.” This toolkit is designed to provide guidance and tools for developing “safe and responsible” AI applications. It reflects Google’s commitment to addressing the ethical considerations associated with AI technology.
Overall, Google’s introduction of Gemma represents a significant development in the field of AI language models. By offering free and open models that can run locally, Google aims to empower developers and promote transparency in AI applications. With Gemma’s superior performance and the accompanying Responsible Generative AI Toolkit, Google is taking important steps towards advancing the responsible use of AI technology.