DeepSeek-R1

DeepSeek-R1Publishing Background

DeepSeek-R1 is a groundbreakingAI inference model, launched in January 2025 by the DeepSeek team. The model was trained through large-scale reinforcement learning to achieve the same level of performance as theOpenAIDeepSeek-R1 is unique in that it does not rely on supervised fine-tuning, but goes directly to the reinforcement learning phase and develops its reasoning ability through self-evolution in a "trial-and-error" mode.

In addition, DeepSeek-R1 adopts an open-source strategy that allows developers and researchers worldwide to access and use its code and models for free, greatly contributing to the open, transparent and collaborative development of AI technology. The launch of this model marks a major breakthrough in AI's ability to learn and reason on its own.

Open Source Licensing and Model Distillation

• open source license: DeepSeek-R1 is under the MIT License, which is a very liberal open source license. It allows developers to freely use, modify and distribute the model, and commercial use is unrestricted without additional applications.
• Model Distillation: DeepSeek-R1 supportModel Distillation TechnologyIt's a way of puttingLarge ModelThe technique of transferring knowledge to small models can improve the efficiency of small models without losing too much performance. This feature helps developers to train more efficient and targeted small models to promote the application of AI technology in different scenarios.

DeepSeek-R1 Performance

• Benchmarking OpenAI o1: DeepSeek-R1 benchmarks the official OpenAI o1 version in terms of performance, and the key to its success is the large-scale use of reinforcement learning techniques in the post-training phase. The technique significantly improves the model's inference even with very little labeled data.
• Actual test data: DeepSeeker R1 performs well in tasks such as mathematics, coding, and natural language reasoning. For example, in the AIME 2024 (Mathematics Competition) Pass@1 Among the indicators, DeepSeek-R1 reached 96.6%, which is comparable to the official version of OpenAI o1; In the MATH-500 test, DeepSeeker R1 Pass@1 The score is also 94.3%, which is on par with the official version of OpenAI o1.

Miniaturized distillation

• DeepSeek-R1 based miniaturization model: The DeepSeek-R1 team distilled six small models based on this model and made them open source. Among them, the small models of 32B and 70B benchmark OpenAI o1 mini in multiple abilities. For example, DeepSeek-R1-Distill-Qwen-32B was tested in AIME 2024 competition questions Pass@1 Reaching 72.6%, surpassing OpenAI o1 mini's 63.6%; In the MATH-500 test, the former Pass@1 The score is 94.3%, which is also better than OpenAI o1 mini's 90.0%.
• Limitations of Small Models: Although small models perform well in certain tasks, there may be a performance gap compared to large models when facing complex scenarios and large-scale data. For example, when dealing with complex semantic understanding and generation of long texts, the contextual understanding and logical coherence of small models may not be as good as that of large models.

DeepSeek-R1 Application and APIs

• Application ConvenienceDeepSeek-R1 is very easy to use. Users can log on to the official website (chat.deepseek.com) or the official app, turn on the "Deep Thinking" mode, and then call it to deal with a variety of reasoning tasks, such as code writing, content creation and other scenarios.
• API open to the public: The API for DeepSeek-R1 is also open to the public. By setting model='deepseek-reasoner', developers can call the model's API. but it should be noted that the API pricing is $1 per million for input tokens cache hits, $4 per million for misses, and $16 per million for output tokens. For enterprises and developers who need to use it on a large scale, cost is a factor that cannot be ignored.

Other Features and Benefits of DeepSeek-R1

• multilevel reasoning: Unlike traditional AI reasoning, DeepSeek-R1 employs a multi-layer reasoning approach to optimize responses with thought chaining, consensus and search. This process is called Test-time Augmentation (TTA).
• NVIDIA SupportNVIDIA officially defines DeepSeek-R1 as "an open model of state-of-the-art reasoning capabilities". Combined with Microsoft's cloud computing capabilities, DeepSeek-R1 is expected to accelerate the application of AI technology in various industries.
• Domestic AI Search AccessThe Secret Tower AI Search announced the integration of DeepSeek-R1 full-blooded version, realizing the combination of "the strongest reasoning in China + real-time search on the whole network + high-quality knowledge base". This further improves the accuracy and reliability of AI search and enhances the reasoning ability.

Link to paper: https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf
Open source address:https://github.com/deepseek-ai/DeepSeek-R1