DeepSeek-R1 · GitHub Models · GitHub

DeepSeek-R1 stands out at thinking tasks utilizing a step-by-step training process, such as language, thinking, and coding jobs. It features 671B overall criteria with 37B active criteria, and 128k context length.

DeepSeek-R1 builds on the progress of earlier reasoning-focused models that improved efficiency by extending Chain-of-Thought (CoT) thinking. DeepSeek-R1 takes things further by combining reinforcement knowing (RL) with fine-tuning on carefully selected datasets. It developed from an earlier version, DeepSeek-R1-Zero, which relied solely on RL and showed strong reasoning skills but 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 blends reasoning-oriented RL with supervised fine-tuning on curated datasets, leading to a design that achieves modern efficiency on reasoning criteria.

Usage Recommendations

We recommend adhering to the following configurations when using the DeepSeek-R1 series models, consisting of benchmarking, to attain the expected efficiency:

– Avoid including a system prompt; all instructions need to be consisted of within the user timely.
– For mathematical issues, it is suggested to consist of a regulation in your prompt such as: “Please reason step by action, and put your last response within boxed .”.
– When evaluating design performance, it is recommended to perform several tests and average the results.

Additional suggestions

The model’s reasoning output (contained within the tags) might contain more harmful material than the model’s final action. Consider how your application will utilize or display the thinking output; you might want to reduce the thinking output in a production setting.