Calibration-free top-k selector speeds MoE decoding without threshold tuning

A new top-k channel selector for Mixture-of-Experts models removes the need for offline score-threshold calibration while keeping accuracy and latency gains. Previously, thresholds fitted on held-out data were used to limit routed-expert reads, but this varied across inputs and could trigger repetition loops in models such as Gemma-4. The approach instead keeps each token's k highest-scoring channels, requiring no calibration data. In reported comparisons it matched or beat the threshold version in 40 of 43 cases at 13% and 18% budgets. The authors report retaining 98.1% of full-model accuracy at 17.8% of routed-expert reads on three post-trained MoE models, with Triton kernels running MoE blocks 1.60 to 1.90 times faster than optimized vLLM at batch size one. The evaluation is limited to those three checkpoints and Triton kernels, so generalization to other model families and hardware remains unconfirmed.