Google has made a significant announcement in the world of AI with the introduction of Gemma, a new family of AI language models. Gemma models are free and open-weights models that can run locally on desktop or laptop computers. This release marks Google’s first major open large language model (LLM) since the launch of OpenAI’s ChatGPT in 2022, which created a frenzy for AI chatbots.
Gemma models come in two sizes: Gemma 2B, which has 2 billion parameters, and Gemma 7B, which has 7 billion parameters. These models are available in pre-trained and instruction-tuned variants. Parameters in AI refer to values in a neural network that determine the behavior of the 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 LLMs. The name Gemma is derived from the Latin word “gemma,” meaning “precious stone.” While Gemma is Google’s first major open LLM since ChatGPT, the company has previously made contributions to open AI research with developments like the Transformer architecture, TensorFlow, BERT, T5, and JAX.
Smaller open-weights LLMs have been seen as tech demos until recently, as larger models have started to match the performance levels of GPT-3.5. However, experts believe that source-available and open-weights AI models are crucial for ensuring transparency and privacy in chatbots. It’s important to note that Gemma is not considered “open source” in the traditional sense, as it does not have a specific software license with few restrictions attached.
The introduction of Gemma seems to be a strategic move by Google to compete with Meta, which has been releasing open-weights models since February of last year. Meta’s approach contrasts with OpenAI’s GPT-4 Turbo, which is only available through the ChatGPT application and a cloud API and cannot be run locally. Google’s Gemma release is seen as an attempt to attract more developers to its Vertex AI cloud platform, according to a Reuters report.
Although we have not yet used Gemma, Google claims that the 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. Gemma is available today through Kaggle, a machine-learning community platform, and Hugging Face.
In addition to the Gemma release, Google has also introduced a “Responsible Generative AI Toolkit.” This toolkit aims to provide guidance and tools for developing “safe and responsible” AI applications. With the growing concerns surrounding AI ethics, this toolkit is a step towards ensuring responsible AI development.
Google’s introduction of Gemma and the Responsible Generative AI Toolkit showcases the company’s commitment to advancing AI technology while addressing concerns around transparency, privacy, and ethics. As Gemma becomes more widely adopted by developers, it will be interesting to see how it shapes the future of AI language models and their applications.