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 |
Benchmarks12
| Benchmark | Score |
|---|---|
| MATH | 97.3% |
| GPQA Diamond | 71.5% |
| SWE-bench Verified | 49.2% |
| MMLU | 90.8% |
| AIME 2024 | 79.8% |
| LiveCodeBench | 65.9% |
| MMLU-Pro | 84% |
| IFEval | 83.3% |
| Codeforces Rating | 2029 elo |
| HumanEval | 92% |
| AIME 2025 | 74% |
| Humanity's Last Exam | 8.6% |
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