DeepSeek-R1震撼发布：开源推理大模型，性能比肩OpenAI o1！

今天，我们正式发布DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.，并同步开源模型权重，标志着开源AI社区迎来又一里程碑！

开源许可与技术创新

DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.遵循MIT LicenseA permissive open-source software license allowing commercial use and modification.，为用户提供了极大的使用自由度。特别值得关注的是，我们明确允许用户通过蒸馏技术借助R1训练其他模型，这一举措将极大促进AI技术的普及与创新。

API服务全面开放

DeepSeek-R1 APIThe API service for accessing DeepSeek-R1's reasoning capabilities with competitive pricing.已正式上线，向所有用户开放思维链输出功能。只需设置model='deepseek-reasoner'即可调用这一强大的推理能力。同时，DeepSeek官网与App已同步更新上线，为用户提供无缝体验。

性能对标行业标杆

强化学习驱动性能突破

DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.在后训练阶段大规模采用了强化学习技术，在仅有极少标注数据的情况下，显著提升了模型的推理能力。在数学、代码、自然语言推理等核心任务上，性能已比肩OpenAI o1A proprietary reasoning large language model developed by OpenAI for advanced reasoning tasks.正式版。

技术论文完全公开

为推动技术社区的交流与协作，我们将DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.的训练技术全部公开：

论文链接：https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf

模型家族全面开源

双660B模型发布

我们同时开源了DeepSeek-R1-ZeroA variant of DeepSeek-R1 trained directly with reinforcement learning on the base model, skipping traditional supervised fine-tuning (SFT).和DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.两个660B参数规模的模型，为研究社区提供强大的基础模型支持。

蒸馏小模型超越竞品

通过DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.的输出，我们蒸馏了6个小模型并开源给社区。其中32B和70B模型在多项能力上实现了对标OpenAI o1-miniA smaller version of OpenAI's o1 reasoning model used as a performance benchmark.的效果，展现了出色的性能表现。

模型获取

HuggingFace链接：https://huggingface.co/deepseek-ai

开放生态建设

许可证优化

为降低开发者的理解成本，我们的开源仓库（包括模型权重）统一采用标准化、宽松的MIT LicenseA permissive open-source software license allowing commercial use and modification.，实现完全开源，不限制商用，无需申请。

协议更新支持蒸馏

我们已更新线上产品的用户协议，明确允许用户利用模型输出、通过模型蒸馏等方式训练其他模型，进一步促进技术的开源和共享。

使用指南

客户端使用

登录DeepSeek官网或官方App，打开“深度思考”模式，即可调用最新版DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks.完成各类推理任务。

API服务定价

DeepSeek-R1 APIThe API service for accessing DeepSeek-R1's reasoning capabilities with competitive pricing.服务采用极具竞争力的定价策略：

输入tokens：每百万tokens 1元（缓存命中）/ 4元（缓存未命中）
输出tokens：每百万tokens 16元

详细文档

如需了解更多API调用细节，请参考官方文档：
https://api-docs.deepseek.com/zh-cn/guides/thinking_mode

结语

Data Analysis

特性/方面	DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks. 详细信息
开源许可	MIT LicenseA permissive open-source software license allowing commercial use and modification. (完全开源，不限制商用)
核心技术创新	强化学习后训练；明确允许模型蒸馏
模型规模	660B 参数 (R1-Zero 和 R1 两个版本)
对标性能	核心任务性能比肩 OpenAI o1A proprietary reasoning large language model developed by OpenAI for advanced reasoning tasks. 正式版
蒸馏小模型	32B 和 70B 模型对标 OpenAI o1-miniA smaller version of OpenAI's o1 reasoning model used as a performance benchmark.
API 调用模型名	`model='deepseek-reasoner'`
API 定价 (输入)	1元 / 百万 tokens (缓存命中) / 4元 (缓存未命中)
API 定价 (输出)	16元 / 百万 tokens
主要访问方式	官网/App “深度思考”模式；API 服务
技术论文	已完全公开于 GitHub
模型获取	HuggingFace 仓库

Source/Note: 信息综合自提供的 DeepSeek-R1An open-source reasoning large language model developed by DeepSeek that performs comparably to OpenAI o1 on mathematical, coding, and natural language reasoning tasks. 发布文本。