Activated Parameters

Activated parameters refer to the subset of a model’s total parameters that are actively utilized during inference for a specific input, as opposed to the model’s total parameter count. This concept is critical in Mixture-of-Experts (MoE) architectures, where only a fraction of parameters are dynamically activated per token, optimizing computational efficiency without sacrificing performance.

Key Examples

References

  • 2026 04 24 DeepSeek V4 Next Gen Open Source LLM Performance and Efficiency Analysis

Source Notes