C.W.K.
Stream
Lesson 01 of 04 · published

Three Paradigms, One Decision Table

~12 min · compare, decision, tradeoffs

Level 0Scout
0 XP0/41 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete

Putting it all together

You now have three lenses — dense, MoE, reasoning — and four axes — backbone, training, inference, product. The combinations are not infinite, but they are too many to remember as a list. This lesson is the master decision table: for each dimension that actually matters in production, which paradigm wins?

DimensionDenseMoEReasoning-oriented
Per-token computeHighest (all params active)Lowest (only top-K experts)Variable, depends on thinking depth
Memory footprint= total params= ALL params (≫ active)= base model + KV for thinking tokens
Latency (per request)PredictableLower compute but routing overheadHigh and variable (5–20× standard)
Serving complexitySimple — standard tensor parallelismComplex — expert parallelism, balancingModerate — long sequences, budget control
Fine-tuningEasy (mature LoRA ecosystem)Hard (routing complicates adaptation)Very hard (RL pipelines, reward models)
Local deploymentBest (llama.cpp, Ollama)Challenging (huge total memory)OK if base fits, just slower
Cost per token (API)Proportional to sizeLower per-FLOP, but memory amortizationMuch higher due to thinking tokens
Quality on hard reasoningGood at scaleGood (training-dependent)Excellent on hard problems
Quality on simple Q&AExcellent (cheap and fast)Excellent (cheap and fast)Wasteful — overthinks simple questions

The two big patterns to internalize

Pattern 1. Dense and MoE compete on the backbone axis (cost per token vs cost per memory). Reasoning is orthogonal — it can be layered on top of either. So "MoE vs reasoning" is a malformed comparison; the real comparison is "MoE vs dense" and "reasoning on vs off".

Pattern 2. Each row in this table maps to a real production constraint. If your bottleneck is memory, MoE makes it worse; if your bottleneck is compute, MoE helps; if your bottleneck is serving simplicity, dense wins; if your bottleneck is hard-task accuracy, reasoning helps; if your bottleneck is latency, reasoning hurts.

External links

Exercise

Take any LLM-using workload you actually run (chatbot, agent, coding assistant, summarizer). For each row in the table above, mark which side is your bottleneck. Then read down the columns — the paradigm with the most green for your bottlenecks is your match. If you've been using something else, that is a useful signal.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.