It's a dense model, so 27B active parameters to compute. Compare that to DeepSeek V4 Flash, which has only 13B active parameters (MoE).