What Do AI Model Numbers Mean? 3B, 7B, 20B…
When size doesn’t tell the whole story — and what the numbers actually mean
If you’ve ever looked into artificial intelligence, you’ve almost certainly come across expressions like “7B model” or “20 billion parameter LLM”. But what do these numbers actually mean — and more importantly, are they really the most relevant criterion when choosing an AI solution?
Billions of what, exactly?
When an AI model name includes something like “7B” or “70B”, the “B” stands for billion — and it refers precisely to the number of parameters in the model. But what exactly is a parameter?
The simplest analogy is a dial or a tuning knob: a small, adjustable value that the model sets for itself during training. In the human brain, neurons are connected by synapses. Parameters play a similar role in AI. They are not hand-coded by engineers: the model “learns” their optimal values by training on billions of text examples.
What is a parameter, concretely?
Imagine completing a sentence: “The cat sat on the…” The model’s internal mechanism assigns weights to all the possible next words: “mat”, “roof”, “chair”, and so on. These weights — which determine how likely each word is in that position — are parameters.
In a 7 billion parameter model, 7,000,000,000 such weight values are stored. Each one represents a tiny decision about how concepts, words, and grammatical structures connect to each other. Taken together, they form the model’s “knowledge”.
Crucially, parameters do not store words or sentences verbatim. Instead, they encode statistical patterns — how certain concepts, phrases, and relationships co-occur in the training data. The model does not “remember” text; it internalises the structure of language.
How do parameters “learn”?
During training, the model repeatedly attempts to predict the next word in a sequence. When it makes a mistake, a feedback mechanism called backpropagation makes tiny adjustments to the parameters — billions of times, in sequence. The result is a network capable of generating responses that resemble human-written text.
The more parameters, the more such connections can theoretically be stored. In practice, the picture is considerably more nuanced.
| Model size | Parameter count | Typical RAM needed | Common use case |
| 3B | 3,000,000,000 | ~4 GB | Simple tasks, mobile devices |
| 7–8B | 7–8,000,000,000 | ~8–16 GB | Professional local deployment |
| 13–20B | 13–20,000,000,000 | ~16–32 GB | Balanced performance |
| 70B+ | 70,000,000,000+ | 40–80+ GB | Enterprise cloud solutions |
Why a larger model doesn’t guarantee better answers
The common assumption is that a larger model is a smarter model. This view is an oversimplification — and in practice, it is regularly disproved.
The quality of a response from an AI system depends on three main factors, and the model itself is only one of them:
Data quality: ~45% • Prompt quality: ~28% • Model size: ~22%
Data quality is the most decisive factor. What documents, regulations, and professional content does the system consult when forming its answer? For a legal AI, for example, the accuracy and relevance of case law, statutes, and contract templates matters far more than the parameter count.
Prompt quality is equally significant. How is the question phrased? Does it include context, constraints, examples? A poorly worded question will extract a poor answer even from the best model.
The model itself — including its parameter count — comes third. A well-configured 7B model, supplied with precise professional data and queried with well-formed questions, will regularly outperform a 70B generalist model that is poorly prompted or given no relevant context.
Hardware constraints: the overlooked criterion
Parameter count does not only affect knowledge capacity — it directly determines the hardware required to run the model. This consideration rarely features in marketing materials, but it is decisive for real-world deployment.
Running a 70 billion parameter model typically requires 40 to 80 GB of RAM, optimised for GPU acceleration. That implies enterprise-grade servers or cloud infrastructure, with significant running costs.
A 7–8B model, by contrast, runs comfortably on a desktop mini PC or a capable laptop with 16–32 GB of RAM, on a local network, without internet connectivity — and in regulated sectors (law, accounting, healthcare), that offline capability is a meaningful advantage for data protection and compliance.
This is precisely the choice made by ArkeoAI: prioritise compact, professionally optimised models rather than massive cloud-dependent ones. Not size for its own sake, but precision and data sovereignty.
Not all billions are equal: model architecture
There is a further dimension that rarely gets mentioned: parameter count alone does not reveal how the model is built. Newer architectures — such as Mixture of Experts (MoE) — can be substantially more efficient than conventional models of a similar size.
A MoE-based 20B model, for instance, may only activate a fraction of its parameters at any one time (say, 4–5 billion), but those activated parameters are selected precisely for the task at hand. The result: comparable or better performance than a monolithic 20B model that engages all parameters on every step — at a fraction of the computational cost.
Parameter count is the model’s storage capacity. Architecture is how it uses that capacity. Both matter.
What to look for when choosing a model
When evaluating an AI solution, parameter count should not be the first question. These should be:
| Question | Why it matters |
| What data does the system run on? | Data quality is the primary performance driver |
| How are the prompts constructed? | Question quality weighs as heavily as the model |
| Local or cloud-based deployment? | Privacy, compliance, and cost implications |
| What architecture does it use? | MoE vs. dense: efficiency and resource demands |
| What specific task was it optimised for? | General-purpose vs. domain-specific focus |
More parameters does not automatically mean more relevance. What matters is the fit between the model, the data, and the use case. A well-designed system with a modest model will always outperform a poorly-directed giant. Power lies in precision, not in size.
