Kosmos-1

Training compute

3.5×10²¹ FLOP

Parameters

1.6B

Published

Mar 1, 2023

Kosmos-1 is an AI model developed by Microsoft (United States), first published in March 2023. It works in the multimodal, language and vision domain, on tasks such as visual question answering, image captioning, language modeling/generation and 5 more.

Training it took an estimated 3.5×10²¹ FLOP of compute (estimation method: operation counting). The model has 1,600,000,000 parameters. It was trained on roughly 360B datapoints.

Access: Unreleased. Its weights are not openly released. It is built on top of ViT-G/14 (LiT). Epoch AI rates the confidence of this record as likely.

Full record

Organization: Microsoft
Country of organization: United States
Domain: Multimodal, Language, Vision
Task: Visual question answering, Image captioning, Language modeling/generation, Chat, Question answering, Document classification, Image classification, Visual puzzles
Training compute: 3.5×10²¹ FLOP
Compute estimation method: Operation counting
Parameters: 1,600,000,000
Dataset size: 360B
Model accessibility: Unreleased
Open weights: No
Base model: ViT-G/14 (LiT)
Epoch confidence: Likely