Parameter

A parameter is one of the individual numeric weights inside a neural network that gets adjusted during training and collectively encodes everything the model has "learned" — there is no separate database of facts inside an LLM; all of its knowledge and capability is distributed across billions of these numbers, each representing the strength of a connection between artificial neurons. Parameter count has become the most commonly cited (if imperfect) shorthand for model scale and rough capability: models are described as "7B," "70B," or "trillion-parameter" as a quick signal of size, with the general (though not absolute) trend that more parameters, trained on more high-quality data, tends to produce more capable models — a relationship formalized in AI research as "scaling laws." This matters for SaaS builders in very practical ways: parameter count roughly correlates with inference cost, memory footprint, and latency (a 70B-parameter model needs more GPU memory and compute per token generated than a 7B model, and is typically slower and more expensive to run), which is why model selection is fundamentally a trade-off between capability and cost/speed rather than "always pick the biggest model." A concrete example: Claude's model family spans this trade-off explicitly — Claude Opus (the largest, most capable, most expensive tier) suits complex reasoning and high-stakes tasks; Claude Sonnet (a balanced mid-tier) suits most production workloads; Claude Haiku (the smallest, fastest, cheapest tier) suits high-volume, latency-sensitive tasks like real-time chat classification or simple extraction, where the extra capability of a larger model would be wasted expense. Builders architecting a multi-feature AI product often route different tasks to different parameter-scale models within the same family — a support ticket triage feature might use a Haiku-class model given its simplicity and volume, while a complex contract-analysis feature might justify an Opus-class model given the stakes of getting it right — rather than defaulting every call to the largest available model. Note that raw parameter count alone doesn't fully determine quality — training data quality, architecture refinements, and post-training alignment (RLHF) all matter, which is why a newer, well-trained smaller model can outperform an older, larger one. It's worth being skeptical of parameter count as a marketing number in isolation: a well-trained, well-aligned smaller model frequently outperforms a larger, poorly-trained or poorly-aligned one on real tasks, and post-training techniques (RLHF, fine-tuning, RAG grounding) often move the needle on practical output quality more than raw parameter scale alone. When evaluating models for a specific SaaS feature, builders get a far more useful signal from benchmarking actual output quality on representative production examples than from comparing headline parameter counts between providers.