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.
Pricing
| Type | Price per MTok |
|---|---|
| Input | $0.55 |
| Output | $2.19 |
Benchmarks13
General
1 benchmarkReasoning
3 benchmarksMath
3 benchmarksCoding
4 benchmarksPercentile among tested models:Top 25%50-75%25-50%Bottom 25%
DeepSeek Family1
| Model | Tier | Released | Input $/MTok |
|---|---|---|---|
| DeepSeek V3 | — | 2024-12-26 | $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