EXAONE Deep 32B

Training compute

1.3×10²⁴ FLOP

Parameters

32B

Published

Mar 16, 2025

EXAONE Deep 32B is an AI model developed by LG AI Research (South Korea), first published in March 2025. It works in the language domain, on tasks such as language modeling/generation, question answering, quantitative reasoning and code generation.

Training it took an estimated 1.3×10²⁴ FLOP of compute (estimation method: reported,operation counting,hardware). The model has 32,000,000,000 parameters. It was trained on roughly 12B datapoints. Training ran on 512 NVIDIA H100 SXM5 80GB for about 2.2K hours.

Access: Open weights (non-commercial). Its weights are openly available. It is built on top of EXAONE 3.5 32B. Epoch AI rates the confidence of this record as confident.

Full record

Organization: LG AI Research
Country of organization: South Korea
Domain: Language
Task: Language modeling/generation, Question answering, Quantitative reasoning, Code generation
Training compute: 1.3×10²⁴ FLOP
Compute estimation method: Reported, Operation counting, Hardware
Parameters: 32,000,000,000
Dataset size: 12B
Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 512
Training time: 2,160 h
Training power draw: 703.2 kW
Numerical format: BF16
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Base model: EXAONE 3.5 32B
Epoch confidence: Confident