Mapping use cases to paradigms
Below is a working map. None of these are absolute laws — but each row reflects a strong prior that should be the default unless you have a specific reason to deviate.
| Use Case | Best paradigm | Why |
|---|---|---|
| Conversational chatbot | Dense (or MoE) standard | Predictable cost, fast responses, good enough for conversation |
| Coding autocomplete | Small dense, standard | Latency-critical; reasoning would feel like the IDE froze |
| Coding debugging assistant | Reasoning for hard, dense for autocomplete | Hard debugging benefits from extended thinking; trivial completions don't |
| Multi-step agent workflow | MoE or small dense, standard | Many calls × per-call latency = bad with reasoning; cost adds up too |
| Long-context retrieval | Dense or MoE, standard | Retrieval is fast pattern matching, not deep reasoning |
| Local on-device inference | Small dense + quantization | Best ecosystem support, predictable memory, fits hardware budgets |
| Research/analysis (one-shot) | Reasoning | Multi-step analysis benefits from extended thinking |
| Creative writing | Standard (any backbone) | Reasoning often makes creative output more mechanical, not better |
| Math word problems / proofs | Reasoning | Where the reasoning paradigm shines most cleanly |
| Cost-sensitive production API | MoE standard or small dense | MoE for capability/FLOP, small dense for predictability |
| Frontier-capacity research | MoE + reasoning if hard | Maximum capability, accept the cost |
How to reason past the table
For any new use case, ask three questions:
- Is this a chain of many calls (agentic) or a single hard call (analytic)? Many calls → avoid reasoning. Single hard call → consider reasoning.
- Is the bottleneck latency, cost, or quality? Latency → fast dense. Cost → small dense or MoE. Quality on hard problems → reasoning.
- Will this run locally or on a managed API? Locally → small dense. Managed → any.
The default
If you cannot articulate a specific reason to use anything else: medium dense (Llama 3.3 70B class) in standard inference mode, on a managed API. That is the productive default in 2026.