Google releases new open-source AI models called Gemma plans to take on Meta

Google has made a significant announcement in the realm of artificial intelligence (AI), unveiling a new family of open-source AI models named Gemma.

These models are positioned to compete with offerings from Meta and various well-funded AI startups such as Mistral and Hugging Face. The move marks a departure from Google’s previous stance in favour of proprietary models, acknowledging the growing influence and appeal of open-source alternatives.

Named in homage to Google’s proprietary Gemini models, which are available for a fee, Gemma models represent a strategic shift towards accommodating the preferences of software programmers and engineers.

This shift reflects a broader industry trend where the demand for open-source AI models, despite their historically smaller size and lower capability, has been steadily increasing.

The popularity of open-source models stems from their flexibility and cost-effectiveness, appealing to both developers and companies seeking to manage expenses associated with AI implementation.

Tris Warkentin, Director of Product Management at Google DeepMind, the entity behind Gemma’s development, emphasized the feedback received from programmers who frequently integrate both proprietary and open-source models into their AI applications.

This integration underscores the necessity for a diverse toolkit, leveraging proprietary models for specific high-performance tasks while utilizing open-source alternatives for their customization options.

Google’s decision to offer both proprietary and open-source models aligns with the practical needs of businesses developing AI applications. Consolidating model deployment on a single cloud computing platform streamlines operations and minimizes complexities associated with data transfer between multiple environments.

While Gemma models share foundational principles with Google’s Gemini models, they are tailored for text-only applications initially, unlike Gemini’s multi-modal capabilities. Additionally, Gemma models are designed to be available exclusively in English at launch, contrasting with Gemini’s multi-lingual support.

Addressing concerns about AI safety, Google asserted its commitment to implementing robust safeguards for Gemma. Extensive measures, including meticulous data curation to prevent privacy breaches and rigorous safety testing, have been undertaken. Despite potential risks associated with open-source models, Google emphasized its dedication to responsible use and deployment, offering guidelines and safety filters to mitigate adverse outcomes.

Jeanine Banks, Vice President and General Manager of Developer Relations at Google, emphasized the company’s stringent licensing terms for Gemma, aimed at preventing malicious usage.

Unlike Meta’s restrictive licensing terms, Google opted for a more permissive approach, allowing broader access to Gemma without commercial restrictions.

The Gemma models are available in two sizes, featuring neural networks with 2 billion and 7 billion adjustable parameters respectively, surpassing Google’s smallest proprietary model, Gemini Nano. However, they are likely smaller in scale compared to Gemini’s higher-tier models.

Google’s introduction of Gemma signifies a strategic response to evolving industry dynamics, catering to the growing demand for open-source AI solutions while maintaining a commitment to safety and responsible usage.

(With inputs from agencies)