Google unveils Gemini 1.5, a next-gen AI model with million-token context window

Google today unveiled Gemini 1.5, the latest iteration of its conversational AI system, touting major advances in efficiency, performance and long-form reasoning capabilities.  

The new system, detailed in a blog post by Google AI chief Demis Hassabis, incorporates significant architecture improvements that allow its core model, Gemini 1.5 Pro, to perform on par with the company’s largest Gemini 1.0 Ultra model while using less computing resources. The Gemini 1.0 Ultra model was introduced last week.

However, the biggest leap comes in the form of an experimental million-token context window that Google says represents a “breakthrough in long-context understanding.” The standard Gemini model analyzes prompts within a 128,000 token context. With the million-token upgrade, Gemini 1.5 can process a vastly larger amount of continuous information before generating its response. 

Million-token context allows long-form reasoning

Google CEO Sundar Pichai offered examples of Gemini 1.5’s extended reasoning capacity in his foreword to the announcement, saying the system can now summarize the entire Apollo 11 mission transcript or analyze a silent 44-minute Buster Keaton film in full.

The extended context allows Gemini 1.5 to “seamlessly analyze, classify and summarize large amounts of content within a given prompt,” Hassabis wrote. He said early results show Gemini 1.5 maintains performance even as the context window grows into the millions.

Public availability still unclear

It remains unclear when or if the million-token version will be broadly available. For now, Google is offering a limited preview to developers and enterprise users through its Vertex AI platform.

The release comes just over a week after Google rebranded its conversational AI system from Bard to Gemini and launched a paid Gemini Advanced tier powered by the Ultra 1.0 model. Gemini has been positioned as a rival to OpenAI’s popular ChatGPT Plus system.

Hassabis said the efficiency gains in Gemini 1.5 will help Google’s teams “iterate, train, and deliver more advanced versions of Gemini faster than ever before.”

Pichai noted Google is focused on developing Gemini responsibly in line with its AI principles. The company said Gemini 1.5 underwent extensive ethics and safety testing focused on areas like content safety and representational harms.

The pace of advancement in conversational AI has accelerated dramatically since the launch of ChatGPT late last year. Experts say reduced training costs and innovations like Google’s Sparsely-Gated Mixture-of-Experts architecture are allowing new iterations to be developed far more rapidly than past AI systems.

With Gemini 1.5, Google is signaling it aims to maintain its leadership position in the AI race. The big question now is how long it will take for these powerful long-context reasoning abilities to make their way into Google’s consumer products.