Qwen2.5-72B

Training compute

7.8×10²⁴ FLOP

Parameters

72.7B

Published

Sep 19, 2024

Qwen2.5-72B is an AI model developed by Alibaba (China), first published in September 2024. It works in the language domain, on tasks such as language modeling/generation, question answering and quantitative reasoning.

Training it took an estimated 7.8×10²⁴ FLOP of compute (estimation method: operation counting). The model has 72,700,000,000 parameters. It was trained on roughly 18T datapoints.

Access: Open weights (unrestricted). Its weights are openly available. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Alibaba
Country of organization: China
Domain: Language
Task: Language modeling/generation, Question answering, Quantitative reasoning
Training compute: 7.8×10²⁴ FLOP
Compute estimation method: Operation counting
Parameters: 72,700,000,000
Dataset size: 18T
Numerical format: BF16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	129.1
02	ARC AI2	Challenge score	0.94
03	MMLU	EM	0.85
04	HellaSwag	Overall accuracy	0.85
05	PIQA	Score	0.83
06	WinoGrande	Accuracy	0.82
07	BIG-Bench Hard	Average	0.8
08	TriviaQA	EM	0.72