DeepSeek-R1 · GitHub Models · GitHub

DeepSeek-R1 stands out at reasoning jobs using a step-by-step training process, such as language, clinical thinking, and coding jobs. It includes 671B total criteria with 37B active parameters, and 128k context length.

DeepSeek-R1 builds on the development of earlier reasoning-focused designs that improved performance by extending Chain-of-Thought (CoT) reasoning. DeepSeek-R1 takes things further by combining support knowing (RL) with fine-tuning on thoroughly selected datasets. It progressed from an earlier version, DeepSeek-R1-Zero, which relied entirely on RL and revealed strong reasoning abilities but had problems like hard-to-read outputs and language inconsistencies. To deal with these restrictions, DeepSeek-R1 integrates a small amount of cold-start data and follows a refined training pipeline that mixes reasoning-oriented RL with supervised fine-tuning on curated datasets, leading to a design that attains state-of-the-art efficiency on reasoning benchmarks.

Usage Recommendations

We suggest sticking to the following setups when utilizing the DeepSeek-R1 series models, of benchmarking, to attain the anticipated performance:

– Avoid including a system timely; all guidelines must be consisted of within the user timely.
– For mathematical issues, it is recommended to include an instruction in your prompt such as: “Please factor step by action, and put your last response within boxed .”.
– When assessing model performance, it is suggested to carry out multiple tests and average the outcomes.

Additional recommendations

The model’s thinking output (included within the tags) might contain more damaging material than the model’s final response. Consider how your application will utilize or display the reasoning output; you might wish to suppress the reasoning output in a production setting.