Definition
The parameter count of a machine-learning model is the total number of trainable weights it contains — the individual numbers, learned during training, that encode what the model knows. It is the figure quoted in a model's name: a "7B" model has roughly seven billion parameters, a "70B" model about seventy billion. Parameter count is the single most cited proxy for a model's raw capability, and the main driver of how much memory it needs to run.
Why the number dominates hardware planning
Each parameter must be stored in memory to run the model, so the parameter count, multiplied by the bytes-per-parameter set by its precision, determines the VRAM or RAM required. At 16-bit precision a 7B model needs ~14 GB just for weights; quantized to 4-bit it drops to roughly 4 GB. This is why parameter count and quantization together decide whether a given model fits on your hardware, the core question for any self-hoster.
Bigger is not automatically better
More parameters generally means more capability, but the relationship is not linear and architecture matters enormously. A well-trained smaller model can outperform a larger, poorly-trained one, and techniques like knowledge distillation deliberately pack much of a big model's ability into a small one. In mixture-of-experts models, the total parameter count far exceeds the active count used per token, so the headline number overstates the compute required.
Parameter count sets the floor for memory; pair it with quantization and the model's open-weight release to judge what will actually run on your local hardware.
See what runs on your GPU in the GPU–LLM fit dataset.
In Simple Terms
The parameter count of a machine-learning model is the total number of trainable weights it contains — the individual numbers, learned during training, that encode…