DeepSeek-R1 · GitHub Models · GitHub

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

DeepSeek-R1 builds on the progress of earlier reasoning-focused models that improved performance by extending Chain-of-Thought (CoT) thinking. DeepSeek-R1 takes things further by combining support knowing (RL) with fine-tuning on thoroughly chosen datasets. It progressed from an earlier version, DeepSeek-R1-Zero, which relied exclusively on RL and showed strong thinking abilities but had concerns like hard-to-read outputs and language disparities. To resolve these restrictions, DeepSeek-R1 incorporates a little amount of cold-start information and follows a refined training pipeline that blends reasoning-oriented RL with supervised fine-tuning on curated datasets, leading to a model that accomplishes modern performance on thinking standards.

Usage Recommendations

We advise adhering to the following when using the DeepSeek-R1 series designs, consisting of benchmarking, to accomplish the anticipated performance:

– Avoid including a system prompt; all guidelines ought to be contained within the user timely.
– For mathematical problems, it is suggested to include a directive in your timely such as: “Please reason action by step, and put your last response within boxed .”.
– When examining design performance, it is suggested to conduct multiple tests and average the outcomes.

Additional suggestions

The design’s thinking output (contained within the tags) may consist of more damaging content than the design’s last reaction. Consider how your application will utilize or display the thinking output; you might wish to suppress the reasoning output in a production setting.