[Official FAQ] Corrections and follow-up questions welcome.

Mixture-of-Experts models keep many parameters but activate only some experts per token. Marketing “active params” can understate memory: you may still need to load most weights.

Implications:

• Routing quality and backend support strongly affect speed.
• GGUF / vLLM / other stacks differ in MoE maturity—read release notes.
• Benchmark end-to-end; do not buy hardware on active-param slides alone.

Post real tokens/s with model, backend, and GPU if you can.