Know what a result actually tells you.
A useful model decision connects a claim to its source, configuration and measurement method.
Publisher information
A model card explains the publisher's intended use and reported evaluations. It is a primary source, not an independent measurement on your hardware.
Planning estimates
The memory planner calculates model weights from parameter count and precision, then adds a runtime allowance. It does not measure a running model or guarantee compatibility.
Measured results
A meaningful benchmark identifies the model revision, hardware, runtime, quantization, workload and measurement method. Compare equivalent configurations and repeat the test before drawing conclusions.