Choose a model that fits the task and the device budget.
1. Parameters and memory
Model size on disk is not total runtime memory. Working buffers, context state and application resources add overhead. Measure resident memory rather than estimating only from the download size.
2. Quantization tradeoffs
Lower precision can reduce storage and improve execution, but accuracy and supported acceleration can change. Evaluate the exact quantized artifact on your task before adopting it.
3. License and provenance
Record the artifact origin, version, redistribution terms and integrity hash. A model that can be downloaded is not automatically suitable for bundling in a commercial app.
Worked scenario
A smaller quantized model meets the storage budget but misclassifies critical documents. Its lower latency does not compensate for the failed quality requirement.
Apply it
Build a comparison table for two candidate artifacts with quality, memory, storage and compatibility evidence.
Check your understanding
The selected artifact fits both the device budget and the task acceptance criteria. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.