Google has started external testing of Gemini, its multimodal AI model intended to compete with OpenAI’s GPT-4. A small group of selected companies has access to a stripped-down chat version, while Google continues developing its largest version internally.
A limited test ahead of a wider release
The account comes from three anonymous sources cited by The Information, who said they had direct knowledge of the matter. The test suggests Google is beginning to expose Gemini to customers, though it does not establish when a public launch will happen.
Google had announced Gemini for release this year, and earlier reports pointed to a fall launch. The external testing may indicate that a release is getting closer, but the source does not confirm a launch date.
For now, the version companies can try is described as a reduced chat offering. The most capable Gemini model is still being built inside Google, so the test provides only a partial view of the system the company ultimately plans to offer.
One model family for products and businesses
Google intends Gemini to serve both business customers and its own consumer products. Businesses are expected to get access through the cloud, while the model is also planned for services such as the Bard chatbot and AI features in Workspace.
Through Vertex AI, Google plans to make models in different sizes available. The smaller options could handle simpler tasks at lower cost, giving businesses a way to choose a model suited to the work they need done.
The Information describes Gemini as a set of large language models rather than a single model. The models are expected to support uses including chatbots, text summarization, code, and generating new text. The report says it is unclear whether Gemini will use a network of expert models, as the source describes for GPT-4.
Data access and possible capabilities
One tester cited in the report said Google may be able to draw on information from its own products, including Google Search, alongside public web information. That access could help Gemini interpret what users want and could reduce incorrect answers, according to the tester. These are reported expectations, not confirmed performance results.
The report also says Gemini’s code generation is considered strong enough that Google hopes to compete with Microsoft’s GitHub Copilot. Other capabilities under discussion include analyzing graphs, interpreting data, and using voice commands to take actions on a computer, such as in a browser.
These possibilities span different kinds of work: producing and summarizing text, helping with code, interpreting visual information, and carrying out actions. The source does not say which will be available at launch, or whether every Gemini model will support them.
What is known about Gemini so far
Google officially announced Gemini in May. Early rumors cited in the report suggested it would have at least a trillion parameters, and training was expected to use tens of thousands of Google’s TPU AI chips. Those details are presented as early reports and expectations, rather than as specifications confirmed in the article.
Demis Hassabis, Gemini’s lead manager, said in late June that the model would combine some strengths of the AlphaGo system with the language capabilities of large models. That description points to Google’s ambition for Gemini, while the current external test remains limited to a chat version for selected companies.
The next important signs will be what Google makes available through Vertex AI and its consumer products, and how the different model sizes perform in practice. Until the company releases more information, claims about Gemini’s advantages and planned features remain based on the reported test and expectations outlined so far.