Find the active-parameter mistake in this Mixtral 8×7B claim.

The claim says Mixtral activates all eight 7B experts on every token, giving 56 billion active parameters. That arithmetic treats total expert capacity as per-token compute, which is wrong. Mixtral 8×7B means eight routed expert FFNs, each roughly 7B-class in size, with top-2 routing. Each token runs two expert FFNs plus shared layers (attention, embeddings, norms). Active parameters are roughly 12–13B class, not 56B. Total stored parameters include all eight experts because serving loads the full bank into memory. Active FLOPs and active params scale with k, not N.

The error is conflating total expert capacity with per-token active compute. Mixtral 8×7B names eight experts of ~7B-class FFN size with top-2 routing, not 8×7=56B active per token. Per token, only k=2 expert FFNs execute alongside always-on shared layers. Active params land around 12–13B class. Total params include all eight experts plus shared weights (~47B stored), but memory footprint and active compute are different metrics. In interviews, always separate total params (capacity in memory) from active params (what runs per forward pass). Pricing, latency, and FLOP estimates use the active count.