Skip to content
Longterm Wiki

DeepSeek R1

DeepSeekOpen Weight

DeepSeek R1 was released January 20, 2025 as an open-weight reasoning model that rivaled OpenAI's o1. Built on DeepSeek V3 with reinforcement learning for chain-of-thought reasoning. Scored 97.3% on MATH, 71.5% on GPQA Diamond, and 49.2% on SWE-bench Verified. Also released smaller distilled versions (1.5B to 70B). Released under MIT license, making it the first open-weight model competitive with proprietary reasoning models.

Developer
DeepSeek
Released
2025-01-20
Context Window
128K tokens

Pricing

TypePrice per MTok
Input$0.55
Output$2.19

Benchmarks13

General

1 benchmark
83.3%#8/10
25th percentile

Knowledge

2 benchmarks
84%#1/15
97th percentile
90.8%#5/37
86th percentile

Reasoning

3 benchmarks
71.5%#13/34
63th percentile
50th percentile
92.2%#1/1
50th percentile

Math

3 benchmarks
97.3%#4/31
87th percentile
79.8%#9/12
29th percentile
74%#6/7
21th percentile

Coding

4 benchmarks
92%#6/25
74th percentile
65.9%#4/9
61th percentile
38th percentile
19th percentile
Percentile among tested models:Top 25%50-75%25-50%Bottom 25%

DeepSeek Family1

ModelTierReleasedInput $/MTok
DeepSeek V32024-12-26$0.27

Details

Model FamilyDeepSeek
Generationr1
Release Date2025-01-20
Parameters671B
Context Window128K tokens
Open WeightYes
Modalitytext

Capabilities2

reasoningtool-use

Sources1

Tags

deepseekreasoningopen-weightfrontier-ai