The trained weights inside a model, counted in billions (B). More parameters generally means more capability — and more cost and latency to run.
A model's parameters are the numbers it learned during training — the knobs that encode everything it knows. Count is quoted in billions: an 8B model has 8 billion weights, a 70B model has 70 billion. It's the closest thing to a model's "size."
More parameters usually means more capability, but also more memory, higher price and slower inference. That's why labs ship tiers: a small 8B model for cheap high-volume work, a large flagship for the hardest tasks. The frontier closed models (GPT, Claude, Gemini) don't disclose parameter counts at all — so for them, "small" means the efficient tier (Haiku, mini, Flash), not a public number.
Don't over-index on raw size. Architecture, training data and post-training matter just as much: a well-trained 30B model routinely beats an older 70B one. Judge by benchmarks and price-performance, not parameter count alone.