ModelDeepSeek V2 Lite

DeepSeek-V2-Lite

A lightweight MoE large language model version of DeepSeek-V2, with approximately 16B total parameters and approximately 2.4B activated parameters, using MLA attention and a fine-grained expert design.

Timeline (last 90 days)

Nov 2

The study validated on two models, DeepSeek-V2-Lite and Qwen3-30B-A3B.

Unverified50%

Nov 2

DeepSeek-V2-Lite is the lightweight MoE version of DeepSeek-V2, with about 16B total parameters and about 2.4B activated parameters, adopting MLA attention and a fine-grained expert design.

Unverified50%

All Facts (2)

Unverified

DeepSeek-V2-Lite is the lightweight MoE version of DeepSeek-V2, with about 16B total parameters and about 2.4B activated parameters, adopting MLA attention and a fine-grained expert design.

50%Unverified

The study validated on two models, DeepSeek-V2-Lite and Qwen3-30B-A3B.

50%

Source Articles

A New Perspective on Expert Pruning: Structural Redundancy and Preservation Framework for Large MoE ModelsNov 2