Large Open-Weight MoEs Are Converging on Lower Expert Activation Ratios

Analysis of six large open-weight mixture-of-experts (MoE) models from 2025 and 2026 shows the average routed-expert activation ratio dropped from 3.7% to 2.1%, indicating a trend toward larger expert pools with relatively flat top-k values and more specialized experts without proportional compute growth.

Recent large open-weight MoEs are trending toward lower routed-expert activation ratios. Across six of the largest open-weight MoEs released in each year, the average ratio falls from roughly 3.7% in 2025 to 2.1% in 2026. 2026: Kimi K3: 16 / 896 DeepSeek V4 Pro: 6 / 384 LongCat-2.0: 12 / 768 Kimi K2.7: 8 / 384 Inkling: 6 / 256 GLM-5.2: 8 / 256 2025: Kimi K2 Thinking: 8 / 384 DeepSeek V3.2: 8 / 256 Qwen3-Coder-480B: 8 / 160 MiniMax M1: 2 / 32 Llama 4 Maverick: 1 / 128 GLM-4.6: 8 / 160 The trend: larger expert pools, relatively flat top-k, and potentially more specialized experts without proportional growth in per-token compute.