{"slug": "large-open-weight-moes-are-converging-on-lower-expert-activation-ratios", "title": "Large Open-Weight MoEs Are Converging on Lower Expert Activation Ratios", "summary": "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.", "body_md": "Recent large open-weight MoEs are trending toward lower routed-expert activation ratios.\nAcross 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.\n2026:\nKimi K3: 16 / 896\nDeepSeek V4 Pro: 6 / 384\nLongCat-2.0: 12 / 768\nKimi K2.7: 8 / 384\nInkling: 6 / 256\nGLM-5.2: 8 / 256\n2025:\nKimi K2 Thinking: 8 / 384\nDeepSeek V3.2: 8 / 256\nQwen3-Coder-480B: 8 / 160\nMiniMax M1: 2 / 32\nLlama 4 Maverick: 1 / 128\nGLM-4.6: 8 / 160\nThe trend: larger expert pools, relatively flat top-k, and potentially more specialized experts without proportional growth in per-token compute.", "url": "https://wpnews.pro/news/large-open-weight-moes-are-converging-on-lower-expert-activation-ratios", "canonical_source": "https://twitter.com/ayushpai0/status/2077817066619351456", "published_at": "2026-07-16 18:42:14+00:00", "updated_at": "2026-07-16 18:55:17.400778+00:00", "lang": "en", "topics": ["machine-learning", "large-language-models", "ai-research"], "entities": ["Kimi K3", "DeepSeek V4 Pro", "LongCat-2.0", "Kimi K2.7", "Inkling", "GLM-5.2", "Kimi K2 Thinking", "DeepSeek V3.2"], "alternates": {"html": "https://wpnews.pro/news/large-open-weight-moes-are-converging-on-lower-expert-activation-ratios", "markdown": "https://wpnews.pro/news/large-open-weight-moes-are-converging-on-lower-expert-activation-ratios.md", "text": "https://wpnews.pro/news/large-open-weight-moes-are-converging-on-lower-expert-activation-ratios.txt", "jsonld": "https://wpnews.pro/news/large-open-weight-moes-are-converging-on-lower-expert-activation-ratios.jsonld"}}