JOIN OUR WHATSAPP CHANNEL
What is Gemma? Everything to know about Google's release of its new open-source AI models
According to Google, a brand-new family of open models called Gemma helps in the development of AI by researchers and developers.
-
Google has released Gemma 2B and 7B, a pair of open-source AI models
-
The new models were developed by Google DeepMind and other teams within the company
On Wednesday, Google unveiled Gemma, a new artificial intelligence (AI) model that third-party developers may be able to customize and brand as their own. Meta Platforms and other companies have already taken a similar move and now Google has joined the force. The Google DeepMind team and other internal teams created the new models, building upon the groundbreaking Gemini models that were previously available.
What is Gemma?
The Alphabet subsidiary stated that anyone can create AI software for free using its new Gemma family of open models for individuals and companies. The Gemma family of open models is composed of lightweight, cutting-edge models that are constructed using the same technology and research as the Gemini models. According to the company, it is openly disclosing important technical information, including what are known as model weights.
According to Google, the change would encourage engineers to enhance Google's technology and promote the company's recently successful cloud division. According to the company, the models are optimized for Google Cloud, where new cloud users who use them will receive $300 in credits.
Sizes of Gemma models
There are two sizes for the Gemma models: one has a neural network with 2 billion parameters or adjustable variables, and the other has a neural network with 7 billion parameters.
Both sizes are more in line with the 1.8B- and 3.25B-parameter Gemini Nano models, and far smaller than the largest Gemini model, "Ultra," which is said to be well above a trillion parameters. Even though the Gemini Ultra can handle complex or large-scale queries, it needs expensive servers in data centers.
Along with the model weights that came from pretraining, the Gemma models will be made available to developers on Hugging Face. The code for making inferences and optimizing the models will also be provided by Google. The code or data used for pretraining is not being supplied. A trained version and an already-fine-tuned version with pairs of questions and matching answers are available for both Gemma sizes.
Furthermore, Gemma will come equipped with responsible AI toolkits, since guardrails can be more difficult to implement in open models than in Gemini's more controlled approach. Due to the inherent dangers associated with open models, Google stated that the business decided to implement more thorough red-teaming for Gemma.
Other AI companies have also created lighter-weight versions of their main foundation models, though it's unclear how much demand there is for smaller models like Gemma. The smallest version of Llama 2, Llama 2 7B, was released by Meta last year. Google recently revealed a faster Gemini 1.5, again for corporate customers and engineers. Gemini itself comes in many weights, including Gemini Nano, Gemini Pro, and Gemini Ultra.
ALSO READ: What is Sora AI? Everything to Know About the Text-to-Model Tool From OpenAI CEO Sam Altman
Know more about Google's Gemini model
Latest Articles
