DeepSeek-R1 · GitHub Models · GitHub

DeepSeek-R1 excels at thinking tasks using a detailed training procedure, such as language, scientific reasoning, and coding tasks. It includes 671B total specifications with 37B active parameters, and 128k context length.

DeepSeek-R1 constructs on the development of earlier reasoning-focused models that enhanced performance by extending Chain-of-Thought (CoT) thinking. DeepSeek-R1 takes things further by integrating reinforcement knowing (RL) with fine-tuning on thoroughly picked datasets. It developed from an earlier version, DeepSeek-R1-Zero, which relied entirely on RL and showed strong thinking skills however had problems like hard-to-read outputs and language inconsistencies. To address these limitations, DeepSeek-R1 incorporates a percentage of cold-start data and follows a refined training pipeline that mixes reasoning-oriented RL with monitored fine-tuning on curated datasets, resulting in a design that achieves modern efficiency on thinking standards.

Usage Recommendations

We suggest sticking to the following setups when utilizing the DeepSeek-R1 series designs, including benchmarking, to achieve the expected efficiency:

– Avoid including a system timely; all guidelines should be consisted of within the user prompt.
– For mathematical problems, it is a good idea to consist of an instruction in your timely such as: “Please factor action by action, and put your last answer within boxed .”.
– When assessing design performance, it is advised to conduct several tests and average the outcomes.

Additional recommendations

The output (included within the tags) may consist of more hazardous content than the model’s final reaction. Consider how your application will use or show the thinking output; you may wish to suppress the reasoning output in a production setting.