Home / Glossary / Parameters
What is Parameters?
Parameters are the learned weights of a neural network. osFoundry’s catalog labels every model with its parameter count (e.g. 7B, 70B) so you can match model size to your hardware and quality needs.
Detail
Parameters are the weights adjusted during training: 7B means seven billion of them. Parameter count is the quickest proxy for model capacity, but it does not determine quality alone.
Memory scales with bit width times count. FP16 needs ~2 bytes per parameter (7B ≈ 14 GB); 4-bit quantization needs ~0.5 bytes plus overhead (7B ≈ 4-5 GB). Add KV-cache memory on top at runtime.
As a rough map: 1-4B suits on-device and templated transforms, 7-9B is the everyday-assistant entry point, 13-14B adds headroom, 30-70B pulls clearly ahead on long reasoning chains and knowledge breadth, and larger classes live mostly in the cloud.
Generations matter as much as size: training data quality, distillation, and architecture improvements mean a new 8B routinely beats a previous-generation 30B. Use parameter count to pick a size class for your hardware, then choose within the class by recency and task-specific evaluations.
For MoE models, distinguish total from active parameters: memory must hold the total, but each token only runs through the active subset — estimate speed from active, memory from total.
Frequently asked questions
What is the practical difference between 8B and 70B models?
Knowledge breadth, consistency over long reasoning chains, and multi-step instruction following. For simple rewriting or short summaries the gap is often barely noticeable.
How do I estimate memory from parameter count?
Bit width x count: FP16 needs ~2 bytes per parameter (7B ≈ 14 GB), 4-bit quantization ~0.5 bytes plus overhead (7B ≈ 4-5 GB). Add KV-cache memory on top at runtime.
Do MoE models use all their parameters at once?
No — memory must hold the total parameter count, but each token only runs through the active subset. Estimate speed from active parameters and memory from the total.
Related terms
Related features