Foundations · Foundational · Beginner

Model Parameters

The learned weights inside a model, whose count is the usual shorthand for model size.

What Model Parameters is

Parameter count indicates capacity, not quality. Training data volume and quality, architecture and post-training often matter more, and a smaller well-trained model regularly beats a larger poorly trained one.

How it works

Parameters are initialised randomly and adjusted by gradient descent during training. At inference they must be held in memory, which is why size drives hardware requirements.

Why it matters

Understanding that parameters are memory cost, not intelligence, prevents a lot of poor model selection.

Common uses

→Hardware sizing
→Model comparison
→Quantisation planning

Watch for

✓Widely misread as a quality score

Continue exploring

Scaling Laws Quantization Mixture of Experts Small Language Model