Compare to

Discover how Google's Gemini Ultra and Meta's Llama 3.2 90B stack up against each other in this comprehensive comparison of two leading AI language models.

Released in December 2023 and September 2024 respectively, these models represent significant advancements in artificial intelligence, with Gemini Ultra offering a 32,800-token context window and Llama 3.2 90B featuring a 128,000-token capacity. Their distinct approaches to natural language processing are reflected in their benchmark performances, with Gemini Ultra achieving 83.7% on MMLU and Llama 3.2 90B scoring 86%, making this comparison essential for developers and organizations seeking the right AI solution for their specific needs.

Models Overview

Google Gemini Ultra
Google Llama 3.2 90B

Provider

Company that developed the model
Google Meta

Context Length

Maximum number of tokens the model can process
32.8K 128K

Maximum Output

Maximum number of tokens the model can generate in a single response
8192 Unknown

Release Date

Date when the model was released
06-12-2023 25-09-2024

Knowledge Cutoff

Training data cutoff date
Unknown December 2023

Open Source

Whether the model's code is open-source
FALSE TRUE

API Providers

API providers that offer access to the model
Vertex AI Azure AI, AWS Bedrock, Vertex AI, NVIDIA NIM, IBM watsonx, Hugging Face

Pricing Comparison

Compare the pricing of Google's Gemini Ultra and Meta's Llama 3.2 90B to determine the most cost-effective solution for your AI needs.

Google Gemini Ultra
Google Llama 3.2 90B

Input Cost

Cost per million input tokens
Pricing not available Pricing not available

Output Cost

Cost per million tokens generated
Pricing not available Pricing not available

Comparing Benchmarks and Performance

Compare the performances of Google's Gemini Ultra and Meta's Llama 3.2 90B on industry benchmarks. This section provides a detailed comparison on MMLU, MMMU, HumanEval, MATH and other key benchmarks.

Google Gemini Ultra
Google Llama 3.2 90B

MMLU

Evaluating LLM knowledge acquisition in zero-shot and few-shot settings.
83.7% 86%

MMMU

A wide ranging multi-discipline and multimodal benchmark.
59.4% 60.3%

HellaSwag

A challenging sentence completion benchmark.
Benchmark not available Benchmark not available

GSM8K

Grade-school math problems benchmark.
88.9% Benchmark not available

HumanEval

A benchmark to measure functional correctness for synthesizing programs from docstrings.
74.4% Benchmark not available

MATH

Benchmark performance on Math problems ranging across 5 levels of difficulty and 7 sub-disciplines.
53.2% 68%

Compare More Models