DeepSeek releases DeepSeek-R1 as an open reasoning model

Background

The predecessor, DeepSeek-R1-Zero, was trained with large-scale reinforcement learning without supervised fine-tuning as a first step. It showed problems such as endless repetition, poor readability and language mixing. ^[2]

DeepSeek-R1 added cold-start data before reinforcement learning to address those problems. It has 671 billion total parameters, 37 billion of them activated, and a 128K context length. ^[2]

What happened

On 20 January 2025 DeepSeek released DeepSeek-R1, claiming performance on par with OpenAI o1 on math, code and reasoning tasks. The paper was posted to arXiv on 22 January 2025. ^[1][3]

Key details

DeepSeek reported 79.8% pass@1 on AIME 2024, 97.3% on MATH-500, and the 96.3rd percentile on Codeforces. ^[3]

The code and weights were released under the MIT License, and DeepSeek allowed API outputs to be used for fine-tuning and distillation. ^[1]

DeepSeek also open-sourced six dense models distilled from R1, at 1.5B, 7B, 8B, 14B, 32B and 70B parameters, based on Qwen2.5 and Llama 3 models. It said the 32B and 70B versions were on par with OpenAI o1-mini. ^[1]

The API price at launch was $0.55 per million input tokens on a cache miss, $0.14 on a cache hit and $2.19 per million output tokens, under the model name deepseek-reasoner. ^[1]

A reasoning model is a language model trained to spend more of its output working through intermediate steps before it gives a final answer. Reinforcement learning rewards answers that turn out correct, and that reward is what pushes the model toward longer, checked reasoning. pass@1 measures the share of problems a model solves on its first attempt, without retries or voting.

Impact

On 27 January 2025 Nvidia shares fell 17%, losing close to $600 billion in market value, the largest one-day drop for any company in U.S. history. CNBC linked the sell-off to DeepSeek, whose app had passed ChatGPT as the most-downloaded free app on Apple's U.S. App Store. ^[4]

The R1 paper was later published in Nature (volume 645, pages 633–638, 2025). ^[3]