Choosing models by workload
Start with the job to be done, not the leaderboard. A practical framework for narrowing the field.
There is no single best model. The right choice depends on what you need the model to do, what inputs it receives, and the constraints your product operates under.
Describe the workload
- What are the inputs? Text, images, audio, documents?
- What output do you need? Free text, structured data, an image, speech?
- How long are typical inputs, and how much context must the model see at once?
- Does the model need to call tools or external APIs?
Match categories to tasks
| Workload | Start with |
|---|---|
| Classification, short replies | Compact language models |
| Multi-step analysis | Reasoning models |
| Code review and generation | Coding-focused models |
| Screenshots and scans | Vision-capable models |
| Semantic search | Embeddings plus reranking |
Evaluate on your own examples
Public benchmarks are a starting point, but your data is what matters. Assemble a small set of representative inputs with expected outputs and compare two or three candidates side by side.
Revisit the choice
Models evolve and workloads change. Recheck the selection when you add a new feature, see quality drift, or when usage volume changes the cost picture.