Dense Transformer Architecture Recap: Every Parameter Active for Every Token

~10 min read

In a standard (dense) Transformer, every token passes through the SAME feed-forward network at every layer — all parameters are active for every single token, which is exactly the property MoE changes.

Dense Transformer Architecture Recap: Every Parameter Active for Every Token is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • In a dense Transformer, every token passes through the SAME complete feed-forward network at every layer
  • 100% of the FFN's parameters are active for 100% of tokens, every single time — no selection or routing involved
  • This means parameter count and compute-per-token scale together in lockstep — a bigger dense model costs proportionally more compute per token
  • This coupling is exactly the scaling problem: you can't grow capacity without growing per-token cost proportionally in a dense architecture
  • MoE's entire value proposition is breaking exactly this coupling — the focus of the next subtopic