Aria

Training compute

1.4×10²³ FLOP

Parameters

24.9B

Published

Oct 8, 2024

Aria is an AI model developed by Rhymes AI (United States), first published in October 2024. It works in the multimodal, language, video and vision domain, on tasks such as language modeling/generation, visual question answering, image captioning and 4 more.

Training it took an estimated 1.4×10²³ FLOP of compute (estimation method: operation counting). The model has 24,900,000,000 parameters. It was trained on roughly 6.8T datapoints.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 130 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Rhymes AI
Country of organization: United States
Domain: Multimodal, Language, Video, Vision
Task: Language modeling/generation, Visual question answering, Image captioning, Question answering, Code generation, Video description, Character recognition (OCR)
Training compute: 1.4×10²³ FLOP
Compute estimation method: Operation counting
Parameters: 24,900,000,000
Dataset size: 6.8T
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 130
Epoch confidence: Confident

Benchmark results

01

Video-MME

Overall (no subtitles)

0.68