Google is launching Gemma 2B and 7B, two open-source artificial intelligence models that allow developers to more broadly leverage the research conducted at Gemini.
While Gemini is an AI model that directly competes with (and is almost as powerful as) OpenAI 's ChatGPT , Gemma is probably preferable for simpler tasks, such as simple chatbots.
Read also: Google: Renames Bard AI to Gemini and launches new app
However, these models may make up for the complexity with speed and lower cost of use. Google claims that Gemma models significantly outperform larger models on critical benchmarks and can be run directly on a laptop or desktop. They will be available through Kaggle, Hugging Face, Nvidia's NeMo, and Google's Vertex AI platforms.
Gemma's presence in the open-source world is significantly different from the way Gemini was introduced. While developers rely on Gemini, they work either through the API or through Google's Vertex AI platform. By making Gemma an open-source model, more users can explore Google's AI for better access rather than looking for alternatives.
Both sizes of models will be available under a commercial license, regardless of the size of the organization, number of users, and type of project. However, Google — like other companies — often prohibits the use of its models for specific tasks, such as developing weapons programs.
Gemma will also ship with “responsible AI,” as open-source models can be harder to guard against than more closed systems like Gemini. Tris Warkentin, director of product management at Google DeepMind, said the company undertook “more extensive red-teaming with Gemma because of the inherent risks of open models.”
See more: Gemini AI: Google & Samsung team up to bring it to the Galaxy S24
The AI tools are responsible for allowing developers to create their own guidelines or a list of banned words when deploying Gemma in their projects. In addition, it includes a model debugger that allows users to examine Gemma's behavior and fix problems.
According to Warkentin, for now the models perform best on tasks related to the English language. “We hope to work with the community to meet market needs beyond the English language,” he told reporters.
Developers can use Gemma for free on Kaggle, while first-time Google Cloud users are charged $300 (~€278). The company says researchers can apply for up to $500,000 (~€463,000) in cloud fees.
While it's unclear how much demand there is for smaller models like Gemma, other AI companies have released smaller versions of their flagships. Meta last year released the Llama 2 7B, a smaller version of the Llama 2. Gemini comes in several versions, including the Gemini Nano, Gemini Pro, and Gemini Ultra, while Google recently announced a faster Gemini 1.5 — aimed primarily at professional users and developers.
See also: Google's Gemini is available through Vertex AI
Source: theverge.com
