Mixture of experts (MoE)
A model architecture where each token is routed to only a few of many "expert" sub-networks. Total parameters are huge, but the compute per token is small - which is why these models are cheap to serve.
What it is
A dense model runs every parameter for every token. An MoE model has, say, 128 expert blocks per layer but a router picks 2-8 of them per token. So a model with 400B total parameters might only use 30B active parameters on any given token.
You still need enough vram to hold all the experts (total params), but the speed and cost behave like the active-param count.
Why it matters
MoE is why several of the cheapest capable models on the pricing table (DeepSeek, Mixtral, Qwen, GPT-oss, Llama 4) undercut dense models of similar quality. When you self-host one, size the GPU by total params for memory but expect throughput closer to the active-param tier.