DeepSeek-R1 · GitHub Models · GitHub

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

DeepSeek-R1 develops on the development of earlier reasoning-focused models that improved efficiency by extending Chain-of-Thought (CoT) reasoning. DeepSeek-R1 takes things even more by combining support knowing (RL) with fine-tuning on carefully chosen datasets. It developed from an earlier variation, DeepSeek-R1-Zero, which relied exclusively on RL and showed strong reasoning abilities however had issues like hard-to-read outputs and language disparities. To attend to these constraints, DeepSeek-R1 integrates a percentage of cold-start data and follows a refined training pipeline that blends reasoning-oriented RL with monitored on curated datasets, resulting in a design that accomplishes state-of-the-art efficiency on thinking benchmarks.

Usage Recommendations

We suggest adhering to the following configurations when using the DeepSeek-R1 series designs, including benchmarking, to attain the expected performance:

– Avoid including a system prompt; all instructions must be consisted of within the user timely.
– For mathematical issues, it is a good idea to include a regulation in your timely such as: “Please factor step by step, and put your final answer within boxed .”.
– When assessing design performance, it is advised to carry out multiple tests and balance the outcomes.

Additional recommendations

The model’s thinking output (contained within the tags) might contain more damaging content than the model’s last action. Consider how your application will utilize or show the reasoning output; you might wish to reduce the thinking output in a production setting.