DeepSeek V3

DeepSeekOpen Weight

DeepSeek V3 was released December 26, 2024. A 671B parameter mixture-of-experts model (37B active per token). Achieved GPT-4o-level performance at a fraction of the training cost — reportedly trained for just \$5.6M using FP8 mixed precision on 2,048 H800 GPUs. Scored 88.5% on MMLU and 90.2% on MATH. Released under MIT license. API pricing at \$0.27/\$1.10 per million tokens — one of the cheapest frontier-class models.

Released

2024-12-26

Context Window

128K tokens

Pricing

Type	Price per MTok
Input	$0.27
Output	$1.10

Benchmarks12

General

1 benchmark

86.1%#6/10

45th percentile

Knowledge

2 benchmarks

88.5%#18/37

53th percentile

75.9%#9/15

43th percentile

Reasoning

2 benchmarks

87.6%#3/8

69th percentile

59.1%#19/34

46th percentile

Math

3 benchmarks

90.2%#10/31

69th percentile

89.3%#2/2

25th percentile

39.2%#12/12

4th percentile

Coding

4 benchmarks

82.6%#18/25

30th percentile

40.5%#8/9

17th percentile

Codeforces Rating

1,759#4/4

13th percentile

SWE-bench Verified

42%#17/18

8th percentile

Percentile among tested models:Top 25%50-75%25-50%Bottom 25%

DeepSeek Family1

Model	Tier	Released	Input $/MTok
DeepSeek R1	—	2025-01-20	$0.55

Details

Model FamilyDeepSeek

Generation3

Release Date2024-12-26

Parameters671B

Context Window128K tokens

Open WeightYes

Modalitytext

Capabilities1

tool-use

Sources1

DeepSeek-V3 Technical Report

Tags

deepseekmixture-of-expertsopen-weightcost-efficient