DeepSeek-R1

Training compute

3.5×10²⁴ FLOP

Parameters

671B

Published

Jan 20, 2025

DeepSeek-R1 is an AI model developed by DeepSeek (China), first published in January 2025. It works in the language domain, on tasks such as language modeling/generation, code generation, quantitative reasoning and question answering.

Training it took an estimated 3.5×10²⁴ FLOP of compute (estimation method: operation counting). The model has 671,000,000,000 parameters. It was trained on roughly 14.8T datapoints. The compute alone is estimated at $7M in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. It is built on top of DeepSeek-V3. Epoch AI rates the confidence of this record as confident.

Full record

Organization: DeepSeek
Country of organization: China
Domain: Language
Task: Language modeling/generation, Code generation, Quantitative reasoning, Question answering
Training compute: 3.5×10²⁴ FLOP
Compute estimation method: Operation counting
Parameters: 671,000,000,000
Dataset size: 14.8T
Training cost (2023 USD): $7M
Numerical format: FP8
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Base model: DeepSeek-V3
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	139.7
02	LiveBench	Global average	71.6
03	ForecastBench	Overall score	60
04	Aider Polyglot	Percent correct	56.9
05	Lech Mazur Writing	Mean score	8.3
06	Algotune	Score	1.7
07	MATH Level 5	mean_score	0.93
08	GPQA Diamond	mean_score	0.69
09	OTIS Mock AIME 2024-2025	mean_score	0.53
10	METR Time Horizons	average_score	0.52
11	WeirdML	Accuracy	0.36
12	SciCode	Score	0.36
13	Balrog	Average progress	0.35
14	Fictionlivebench	120k token score	0.33
15	SimpleBench	Score (AVG@5)	0.31
16	ARC-AGI	Score	0.16
17	ARC-AGI-2	Score	0.01
18	Critpt	Accuracy	0.01