Gemini Ultra

Definition

Gemini Ultra is the largest and most capable model in Google DeepMind's original Gemini 1.0 family, announced in December 2023 and made broadly available in February 2024 through the Gemini Advanced subscription, notable at launch for being the first model reported to exceed human-expert-level performance on the MMLU benchmark.

Overview

Gemini Ultra was introduced alongside Gemini Pro and Gemini Nano as the top tier of Google DeepMind's first Gemini model family, positioned as Google's answer to GPT-4 following a period of intense competitive pressure in the large-model race. Google DeepMind reported that Gemini Ultra achieved a score above 90% on the Massive Multitask Language Understanding (MMLU) benchmark, a widely used evaluation spanning 57 academic subjects, describing it as the first model to exceed the approximate threshold associated with human-expert performance on that benchmark, alongside strong results on numerous other reasoning and multimodal evaluations. Like the rest of the Gemini 1.0 family, Ultra was designed from the ground up as natively multimodal, trained jointly on text, images, audio, and code rather than combining separate single-modality models, which Google DeepMind argued gave it stronger cross-modal reasoning than approaches that bolt vision or audio capability onto a text-only model after the fact. Following its initial announcement in December 2023, Gemini Ultra became broadly available to consumers in February 2024 as the engine behind Gemini Advanced, a premium subscription tier of Google's Gemini chatbot app (which itself replaced the earlier Bard branding), offering capabilities beyond the free tier powered by the smaller Gemini Pro model. Enterprise and developer access to Gemini Ultra was also made available through Google Cloud's Vertex AI platform. Gemini Ultra was subsequently succeeded within Google's lineup by the Gemini 1.5 and later Gemini 2.x model generations, which introduced major advances such as the long-context breakthroughs of 1.5, and Google's public branding and product naming shifted accordingly as newer, more capable models replaced Ultra as the flagship offering.

Key Features

• Largest, most capable model in the original Gemini 1.0 family (alongside Pro and Nano)
• Reported to exceed the human-expert performance threshold on the MMLU benchmark at launch
• Natively multimodal, jointly trained on text, images, audio, and code
• Powered the premium Gemini Advanced consumer subscription tier
• Announced December 2023, broadly available from February 2024
• Available to enterprises and developers via Google Cloud Vertex AI
• Positioned as Google's direct competitive answer to GPT-4
• Later succeeded by the Gemini 1.5 and Gemini 2.x model generations

Use Cases

Alternatives

Frequently Asked Questions

Related in AI Models