技术规格

系统要求

不同量化方法和上下文大小的显存要求

DeepSeek-R1 671B

DeepSeek-R1 represents a class of advanced reasoning models developed by DeepSeek, designed to facilitate complex computational tasks and logical inference. It is built upon a Mixture-of-Experts (MoE) architecture, featuring a total of 671 billion parameters, with approximately 37 billion parameters actively engaged during each inference pass. This architecture, inherited from the DeepSeek-V3 base model, incorporates Multi-head Latent Attention (MLA) for efficient processing of extensive datasets and includes an auxiliary-loss-free strategy for effective load balancing during training. The model further leverages Multi-Token Prediction (MTP) to enhance predictive accuracy and expedite output generation.

The training methodology for DeepSeek-R1 emphasizes reinforcement learning (RL) to cultivate sophisticated reasoning capabilities. Initially, a precursor, DeepSeek-R1-Zero, demonstrated emergent reasoning behaviors such as self-verification and the generation of multi-step chain-of-thought (CoT) sequences through large-scale RL without preliminary supervised fine-tuning (SFT). DeepSeek-R1 refines this approach by integrating a small amount of 'cold-start' data prior to the RL stages, which addresses challenges observed in DeepSeek-R1-Zero, such as repetitive outputs and language mixing, thereby enhancing model stability and overall reasoning performance. The training pipeline for DeepSeek-R1 specifically incorporates two RL stages focused on discovering improved reasoning patterns and aligning with human preferences, alongside two SFT stages that initialize the model's reasoning and non-reasoning capabilities.

DeepSeek-R1 is engineered to excel in domains requiring analytical thought, including high-level mathematics, programming, and scientific inquiry. Its design supports a large context length, enabling processing of extended inputs. To broaden accessibility and deployment options, DeepSeek has also released several distilled versions of DeepSeek-R1, ranging from 1.5 billion to 70 billion parameters. These smaller models are designed to retain a significant portion of the reasoning capacity of the full model, making them suitable for environments with more constrained computational resources.