Research / How Chinese Models Use Mixture of Experts

How Chinese Models Use Mixture of Experts

2026-09-22 · updated 2026-09-22

Research question: How do Chinese frontier labs design Mixture-of-Experts architectures in 2026, and what do activation ratios and linear-attention hybrids reveal about the economics of serving these models?

How Chinese Models Use Mixture of Experts — Image: AI-generated illustration (Seedream)

Related entities

deepseek-v4-1-flash
deepseek-v4-pro
kimi-k3
glm-5.3
glm-5.3-flash
minimax-m3
qwen3.8-max
qwen3.8-2.4t-a95b

Sources

[Official] DeepSeek API docs — Models & Pricing [Official] Z.ai docs — GLM-5.3-Flash model page [Official] Z.ai docs — GLM-5.3 model page [Official] Kimi — Kimi K3 release announcement · 2026-07-16 [Official] QwenCloud — Qwen3.8-Max-0902 model page [Official] MiniMax API — model documentation [Official] GitHub — zai-org/GLM-5