Google releases public preview of Gemini 1.5 Pro with one million tokens – Computer – News

Google has released a public preview of Gemini 1.5 Pro. The model has the capacity to input up to a million tokens, far more than AI models to date. Google previously announced the model.

Gemini 1.5 Pro is available within the Vertex AI business service, says Google at its Cloud Next conference. This allows business customers to get started with the model. The company does not say when consumers can start using it.

The version is able to promptly process videos of 1 hour, code files of 30,000 lines of code, or texts of more than 700,000 words in one go. Google is also working on a version that can process ten million tokens at once. This is possible via Vertex AI: the standard version of Gemini 1.5 Pro has a limit of 128,000 tokens.

Gemini 1.5 Pro should also reach approximately the level of Gemini 1.0 Ultra, the most extensive model of the previous generation. Google announced this in December. Another novelty is the use of Mixture-of-Experts architecture. This consists of several smaller models, each of which is an ‘expert’ in its own field. Depending on the prompt, the language model activates one or more such experts to answer the prompt. As a result, for many prompts it is more economical to process the prompt than if the entire language model has to be used for each prompt.