DeepSeek-R1

DeepSeek-R1What is Same

DeepSeek-R1 is a series of reasoning models launched by DeepSeek in 2025, designed for mathematics, coding, and other tasks that require multi-step reasoning. It operates on a different level than the DeepSeek Web Assistant: while the model weights and technical methods are intended for research and deployment, the web product includes a standalone service portal and runtime configuration.

The Difference Between R1 and R1-Zero

R1-Zero explores the development of reasoning capabilities directly through large-scale reinforcement learning, without using supervised fine-tuning as a preliminary step; the authors also note that it suffers from issues such as repetition, readability, and mixed language. R1 builds on this foundation by employing different training procedures to improve output quality; therefore, the training description for R1-Zero cannot be directly applied to R1.

Distillation Models and On-Premises Deployment

Officials have also released distillation models based on the Qwen and Llama architectures, which transfer inference capabilities to smaller-scale models. When selecting a model, one should consider task quality, GPU memory, and inference speed, rather than simply comparing the number of parameters. For more on this approach, seeIntroduction to Knowledge Distillation.

R1 Warehouse and its weights are licensed under the MIT License; however, the distillation model is also subject to the licensing terms of its underlying models. For commercial deployment, you must review the specific model files and documentation used.

Research and Evaluation Materials

Official RepositoryModel documentation and usage materials are provided; for training and evaluation methods, seeTechnical Papers. The results presented in the paper correspond to specific models and test settings; to reproduce them, you must control the sampling, prompts, and evaluation methods simultaneously.