Core Viewpoint - The article discusses the evolution of DeepSeek in the context of Mixture-of-Experts (MoE) models, highlighting innovations and improvements from DeepSeekMoE (V1) to DeepSeek V3, while maintaining a focus on the MoE technology route [1]. Summary by Sections 1. Development History of MoE - MoE was first introduced in 1991 with the paper "Adaptive Mixtures of Local Experts," and its framework has remained consistent over the years [2]. - Google has been a key player in the development of MoE, particularly with the release of "GShard" in 2020, which scaled models to 600 billion parameters [5]. 2. DeepSeek's Work 2.1. DeepSeek-MoE (V1) - DeepSeek V1 was released in January 2024, addressing two main issues: knowledge mixing and redundancy among experts [15]. - The architecture introduced fine-grained expert segmentation and shared expert isolation to enhance specialization and reduce redundancy [16]. 2.2. DeepSeek V2 MoE Upgrade - V2 introduced a device-limited routing mechanism to control communication costs by ensuring that activated experts are distributed across a limited number of devices [28]. - A communication balance loss was added to address potential congestion issues at the receiving end of the communication [29]. 2.3. DeepSeek V3 MoE Upgrade - V3 maintained the fine-grained expert and shared expert designs while upgrading the gating network from Softmax to Sigmoid to improve scoring differentiation among experts [36][38]. - The auxiliary loss for load balancing was eliminated to reduce its negative impact on the main model, replaced by a dynamic bias for load balancing [40]. - A sequence-wise auxiliary loss was introduced to balance token distribution among experts at the sequence level [42]. 3. Summary of DeepSeek's Innovations - The evolution of DeepSeek MoE has focused on balancing general knowledge and specialized knowledge through shared and fine-grained experts, while also addressing load balancing through various auxiliary losses [44].