The one-sentence version
A parameter is a single learned "knob" inside the model — a number the training process tuned. A model's parameter count is the total number of knobs, and it's the single best first-guess at how smart the model is and how much computer it needs.
What Is a Parameter?
During training, a model reads enormous amounts of text and adjusts billions of internal numbers until its predictions stop being wrong. Each of those numbers is a parameter — sometimes called a weight, because it weighs how strongly one piece of information influences another.
Think of a gigantic mixing board with billions of sliders. Training is the process of an engineer moving every slider, one nudge at a time, until the sound comes out right. When training finishes, the sliders stay where they were left. Those frozen positions are the model.
When you download a model, you are downloading the final position of every slider. Nothing else.
What "7B" Actually Means
Model names carry their size: 7B = 7 billion parameters, 70B = 70 billion. Rough intuition for what the tiers buy you:
Tiny. Fast summaries, simple classification, autocompletion. Runs on a laptop or even a phone. Starts struggling with nuance, long documents, and multi-step tasks.
Small but useful. The workhorse tier: solid everyday chat, drafting, document Q&A. Runs comfortably on a single workstation GPU. For many businesses, this is the sweet spot for private internal use.
Mid-size. Genuinely strong reasoning and writing; handles complex instructions well. Needs a serious GPU or careful quantization (see the Quantization page) to run locally.
Frontier class. The best general quality available, and usually a multi-GPU server (or a provider's datacenter). Often a mixture of experts — see the Model Types page.
Bigger Is Not Automatically Better
Parameter count is a capacity measure, not a quality guarantee. Three things break the "more is better" assumption:
Training quality beats size. A modern, well-trained 8B model routinely beats an older 70B one. Newer smaller models are trained on better data with better techniques.
Task shape matters. A 3B model tuned for classification will beat a general 70B at that one job — faster and cheaper.
Speed is a feature. A 70B answer that arrives in 30 seconds is worse than an 8B answer in 2 seconds for many workflows. Latency is part of quality.
Rule of thumb: use the smallest model that does the job well — then step up only when you can name the task it's failing at.
The Hardware Reality
Every parameter must sit in memory (VRAM on a GPU, RAM on a CPU) while the model runs. At standard 16-bit precision that's roughly 2 GB per billion parameters — plus extra for the conversation itself. Ballpark at full precision:
| Model size | Approx. memory (16-bit) | Typical hardware |
|---|---|---|
| 3B | ~6 GB | Laptop CPU / small GPU |
| 8B | ~16 GB | Workstation GPU (or CPU, slowly) |
| 70B | ~140 GB | Multi-GPU server — or quantized down |
That last row is exactly why quantization exists — it's the next page in this track. (Numbers are for intuition, not procurement.)