Google has recently introduced Gemma 2B and 7B, marking a significant shift towards open-source AI models. These models, derived from the flagship Gemini, offer developers greater flexibility in utilizing Google’s research. Despite their compact size, Gemma models exhibit superior performance in key benchmarks, surpassing larger counterparts. An added advantage is their capability to run directly on developers’ laptops or desktops.
Accessible Platforms
Gemma models will be accessible through popular platforms such as Kaggle, Hugging Face, Nvidia’s NeMo, and Google’s Vertex AI. This departure from Gemini’s closed model approach allows developers to experiment with Google’s AI more freely, expanding opportunities beyond the constraints of an API or the Vertex AI platform.
Licensing and Restrictions
Both Gemma 2B and 7B will be available under a commercial license, irrespective of organization size, user count, or project type. However, Google, like other industry players, imposes restrictions on specific applications, such as weapons development programs.
Responsible AI Toolkit
Addressing the challenges of open-source models, Gemma comes with “responsible AI toolkits.” Recognizing the potential risks, Google has conducted extensive testing to enhance safety features. The toolkit empowers developers to establish their recommendations or a list of banned words when implementing Gemma. Additionally, it features a model debug tool for users to scrutinize Gemma’s behavior and troubleshoot issues.
According to Tris Warkentin, director of product management at Google DeepMind, Gemma currently excels in English tasks, but Google aims to collaborate with the community to expand language capabilities.
Usage and Incentives
Developers can leverage Gemma for free on Kaggle, with early Google Cloud users receiving $300 in credits. Researchers can qualify for substantial cloud credits, reaching up to $500,000.
While the popularity of smaller models like Gemma remains uncertain, this release follows a trend in the AI industry, notes NIX Solutions. Meta released Llama 2 7B last year, and Gemini itself offers various iterations, including Nano, Pro, Ultra, and the recently announced Gemini 1.5 for business users and developers.
In conclusion, Google’s Gemma signifies a shift towards open and accessible AI, providing developers with versatile tools for experimentation and development.