Google is already pushing Gemini 1.5, giving its AI a turbo boost

Time really flies, but it has been two months since Google launched its new large language model (LLM) — Gemini. And just a week ago, the company rolled out the high-end version of the AI known as Gemini Ultra. Although Gemini hasn’t been around for all that long yet, Google is already announcing an upgrade to the LLM.

Today, Google revealed it is launching Gemini 1.5 for developers and enterprise users, with a full rollout to consumers in the near future. The first version that will be available to these users is Gemini 1.5 Pro. Gemini has three versions — Nano, Pro, and Ultra — and Pro is the mid-sized general purpose model.

If you’re wondering how 1.5 stacks up to 1.0, Gemini 1.5 Pro reportedly beat Gemini 1.0 Pro on 87% of benchmark tests. Google also claims that Gemini 1.5 Pro operates on almost the same level as Gemini 1.0 Ultra.

In addition to enhanced performance, this upgrade is said to rely on less computational power. This was achieved by using a new Mixture-of-Experts (MoE) architecture, which means it now only uses part of the model when running a query, as opposed to using the entire model. As a result, Gemini 1.5 is faster and more efficient than before.

Another big change coming with this update is the context window. Sundar Pichai highlights that the new model can handle up to 1 million tokens consistently. In comparison, the current version of Gemini Pro only runs 32,000 tokens. If you’re unfamiliar with the meaning of tokens and context windows, you can check out our tokens article, but Google describes them as “building blocks used for processing information.”

As mentioned earlier, Gemini 1.5 Pro will only be available to developers and enterprise users at the moment. These users will be able to access it through Google’s Vertex AI and AI Studio. When it rolls out to consumers, you’ll be able to access the tool on the Gemini website. Google says it will introduce “1.5 Pro with a standard 128,000 token context window when the model is ready for a wider release.” It will also introduce pricing tiers that go above that standard token count, up to 1 million.