Read a memory estimate honestly
Know what is published, calculated, runtime-dependent, or still unknown.
Reviewed 2026-09-08 · Beginner
Published facts and estimates are different
A file’s published byte size is evidence about that artifact. A parameter-count calculation is an estimate of weight storage. Runtime memory adds cache, workspaces, allocation behavior, and other components.
sizeof.ai keeps these distinctions visible. Changing a precision estimate does not create a real downloadable quantization, and a supported context in a model config does not prove acceptable quality at that length.
Three states, three interpretations
An ordinary estimate applies a supported calculation with stated assumptions. A lower bound includes known components but cannot safely account for all runtime-dependent memory. Unavailable means there is not enough trustworthy information for the requested total.
Do not rank an unavailable model as zero GiB or treat a lower bound below capacity as a safe fit. Missing metadata is a reason to inspect the source and measure, not to invent a convenient default.
Freshness and completeness are separate
The search catalog stores discoverable model names and ranking information. Model details are fetched and cached separately. A name appearing in search does not establish runtime support or current artifact completeness.
A cached detail response can be older than its repository. Inspect freshness warnings and source links. The final check is a successful representative workload on the exact hardware and runtime you will use.