DeepSeek-R1

// Overview

DeepSeek-R1 is DeepSeek's first-generation reasoning model, released January 20, 2025. It is a Mixture-of-Experts model built on the DeepSeek-V3-Base — 671 billion total parameters with roughly 37 billion active per token — and it inherits V3's 128K-token context window. DeepSeek released the weights as open weights under the permissive MIT license, which allows commercial use, fine-tuning, and distillation, making R1 the first openly available model to reach reasoning quality DeepSeek described as 'performance on par with OpenAI-o1'.

Unlike a standard chat model, DeepSeek-R1 produces a visible chain-of-thought before its final answer. It was post-trained from V3-Base with a multi-stage pipeline combining reinforcement learning and supervised fine-tuning; the companion model DeepSeek-R1-Zero was trained with pure RL and no SFT cold start, which the team used to show reasoning behavior can emerge from RL alone. The work was published as the paper 'DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning' (arXiv 2501.12948) and later appeared in Nature.

On reasoning benchmarks DeepSeek-R1 scored 79.8 on AIME 2024, 97.3 on MATH-500, 71.5 on GPQA Diamond, and a 2029 Codeforces rating (96.3rd percentile) — figures DeepSeek positioned as comparable to, and on some math and coding tests slightly ahead of, OpenAI's o1. Alongside the flagship, DeepSeek distilled R1's reasoning into six smaller dense models based on Qwen 2.5 and Llama 3 (1.5B to 70B), bringing R1-grade reasoning to commodity hardware. R1's open release and aggressive API pricing (about 90-95% cheaper than o1 at launch) sparked over 500 derivative models on Hugging Face within days.

// Benchmarks

1. AIME 2024 (Pass@1)79.8%
2. MATH-500 (Pass@1)97.3%
3. GPQA Diamond (Pass@1)71.5%
4. LiveCodeBench (Pass@1-COT)65.9%
5. Codeforces Percentile96.3%
6. SWE-bench Verified (Resolved)49.2%
7. Aider-Polyglot53.3%
8. MMLU (Pass@1)90.8%
9. MMLU-Pro (EM)84%
10. MMLU-Redux (EM)92.9%
11. AlpacaEval 2.0 (LC-winrate)87.6%
12. Arena-Hard (vs GPT-4-1106)92.3%
13. FRAMES (Acc.)82.5%
14. IF-Eval (Prompt Strict)83.3%
15. SimpleQA (Correct)30.1%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

DeepSeek's first-party launch pricing for R1 (deepseek-reasoner), effective at the January 20, 2025 release — roughly 90-95% cheaper than OpenAI o1 at the time. That alias has since been remapped to newer models, so the original R1 is no longer the model served at this endpoint. Third-party hosts still serve the original R1 weights; OpenRouter, for example, lists $0.70 in / $2.50 out per 1M tokens.

Pricing source ↗

// Strengths

• Open weights under the permissive MIT license — free for commercial use, self-hosting, fine-tuning, and distillation
• First openly available reasoning model at o1-class quality: AIME 2024 79.8, MATH-500 97.3, GPQA Diamond 71.5
• Strong competition coding — Codeforces rating 2029 (96.3rd percentile), LiveCodeBench 65.9
• Visible chain-of-thought reasoning that effectively self-checks before answering
• Launched roughly 90-95% cheaper per token than OpenAI o1, undercutting closed reasoning models
• Shipped with six distilled dense models (1.5B-70B, Qwen 2.5 / Llama 3) that run on commodity hardware

// Best for

• Competition-style math and multi-step logical reasoning
• Coding and software-engineering tasks (LiveCodeBench, SWE-bench Verified, competitive programming)
• Self-hosted reasoning deployments where an open MIT-licensed model is required
• Distillation: using R1's chain-of-thought traces to train smaller, cheaper student models
• STEM question answering and knowledge tasks (GPQA, MMLU-Pro)
• Research into reinforcement-learning-driven reasoning, building on the open paper and weights

// How to access

// DeepSeek R1 — every version

The full lineage of the DeepSeek R1 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.