Evidence note · reviewed 29 July 2026
Model weights are only the memory floor.
A weight calculation can reject an impossible plan. It cannot, by itself, prove that a model will serve the required context and concurrency.
The serving envelope
01Weightsparameters × precision
→02Runtimekernels and buffers
→03KV cachecontext × users
→04Headroomstability and peaks
→05Testexact engine and load
Total parameters govern resident weights in mixture-of-experts systems even when fewer parameters are active per token. Quantisation changes memory and may change quality or runtime behaviour. Long context and concurrent sessions can dominate the remaining capacity.
Use calculations for planning; use measurements for claims.
- calculate a transparent lower bound from the exact checkpoint
- declare precision, engine, context, batch and concurrency
- reserve explicit headroom rather than filling nominal capacity
- measure quality, latency, throughput and power on the target system
Evidence boundary: calculated memory is not a production benchmark or a warranty of fit.