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-V3 671B

DeepSeek-V3 is a large-scale Mixture-of-Experts (MoE) language model, comprising a total of 671 billion parameters with 37 billion parameters activated per token during inference. This design prioritizes efficient inference and cost-effective training. The model was pre-trained on an extensive dataset of 14.8 trillion diverse and high-quality tokens. Subsequent training phases involved Supervised Fine-Tuning and Reinforcement Learning to further enhance its capabilities. DeepSeek-V3 represents an evolution in large language model design, building on previous architectural foundations while introducing novel advancements for efficiency.

The architectural core of DeepSeek-V3 integrates several innovations. It utilizes Multi-head Latent Attention (MLA), a mechanism designed to optimize attention operations by compressing key-value pairs into a low-dimensional latent space, thereby reducing memory consumption during inference. The Mixture-of-Experts component, termed DeepSeekMoE, employs 256 routed experts and 1 shared expert, with each token dynamically interacting with 8 specialized experts plus the single shared expert. A notable innovation in this MoE architecture is an auxiliary-loss-free strategy for load balancing, which aims to distribute computational load across experts without the performance degradation typically associated with auxiliary loss functions. Additionally, DeepSeek-V3 incorporates a Multi-Token Prediction (MTP) training objective, which densifies training signals and is observed to enhance overall model performance by training the model to predict multiple future tokens simultaneously. Training further leverages FP8 mixed precision, demonstrating its feasibility and effectiveness at an extremely large scale. The model employs Rotary Positional Embedding (RoPE) for handling positional information and RMSNorm for normalization within its layers.

DeepSeek-V3 is engineered to support a broad spectrum of general language tasks, exhibiting capabilities in areas such as mathematical problem-solving, advanced code development, and complex reasoning. Its design allows for the processing of extended contexts, supporting a context length of up to 128K tokens. This enables the model to handle long documents and complex multi-turn conversations effectively. The model's efficiency in both training and inference makes it suitable for applications requiring substantial computational capacity while maintaining resource optimization.