Major Release Google Multimodal Proprietary

Google Gemini Ultra

Google's first model to beat GPT-4 on MMLU — 90%+ with CoT

Released 2023-12-06 · ~540B · knowledge cutoff 2023-06

Overview

Google's most capable multimodal AI model

Specifications

DeveloperGoogle
Release date2023-12-06
Model typeMultimodal
Parameters~540B
ArchitectureMoE Transformer
Knowledge cutoff2023-06
LicenseProprietary
Input modalitiestext, image, audio, video
Output modalitiestext

Benchmarks

Benchmark Gemini Ultra PaLM 2Δ
MMLU 90.0% 78.3% +11.7
HumanEval 74.4% 77.2% -2.8
MATH 53.2% 34.3% +18.9

Availability

Limited Access Google One AI Premium Vertex AI

What's new vs PaLM 2

Direct comparison available in the model comparison tool →

Official references