Google has introduced a collection of no-cost open-source AI models to rival those from Meta and various well-funded AI startups like Mistral and Hugging Face.
These new models, dubbed Gemma, are a tip of the hat to Google’s more advanced Gemini models, which require payment for access.
In the past year, a divide has emerged between advocates of proprietary AI models and proponents of open-source alternatives. While open-source models have historically been smaller and less powerful, the performance gap is narrowing. The flexibility and customizable nature of open-source models have garnered popularity among programmers, engineers, and companies seeking to manage expenses associated with implementing generative AI more effectively.
Google, traditionally aligned with proprietary models alongside OpenAI and Microsoft, is now recognizing the potential advantages of open source, a trend championed by Meta. Amazon’s AWS has also straddled both camps, initially supporting open-source players like Hugging Face before partnering with Anthropic, offering proprietary models.
Tris Warkentin, Google DeepMind’s product management director behind the Gemma models, noted feedback from software developers who often blend proprietary and open-source models in their AI projects. Developers tend to utilize expensive proprietary models only when necessary for specific capabilities, such as generating text descriptions of images.
Businesses leveraging a mix of models and data in their applications benefit from consolidating these resources on a single cloud platform. Google aims to attract more customers to its Cloud Platform by offering both proprietary and open-source models.
While Gemma models share principles with Google’s Gemini models, they are text-based compared to Gemini’s audio, visual, and text input/output capabilities. Additionally, Gemma is initially limited to English, unlike the multilingual Gemini.
Google emphasizes responsibility in AI model safety compared to competitors like OpenAI. Open-source models, while versatile, present challenges in preventing misuse for malicious purposes. Google has implemented robust safeguards in Gemma, including data curation to avoid privacy leaks and extensive safety testing to address potential vulnerabilities.
Gemma is accompanied by guidelines for responsible usage and safety filters to mitigate undesirable outputs. Licensing terms prohibit nefarious use, distinguishing Gemma as an “open model” with usage restrictions, unlike traditional unrestricted open-source software.
Offered in two sizes, Gemma features neural networks with 2 billion and 7 billion adjustable parameters, surpassing the smaller Gemini Nano model with 1.8 billion parameters. However, Gemma is likely smaller than Gemini’s Pro and Ultra models, which boast significantly more parameters, potentially reaching into the tens or hundreds of billions.
