Home/Calculators/Model memory
Parameters × precision

Model weight
memory calculator.

Estimate the memory floor for a custom checkpoint before its full architecture metadata is available.

Custom model calculatorWEIGHT MEMORYFormula v1.0.0
Formulaparameters × bits ÷ 8 × format overhead

4-bit currently uses a 1.16× packing coefficient.

Recommended weight-loading floor10.6 GiBIncludes packing overhead, runtime workspace and 15% operating headroom
Raw weights
6.05 GiB
Packed estimate
7.02 GiB
Runtime workspace
2.00 GiB

This calculator intentionally excludes KV cache because parameter count alone does not reveal layer count, KV heads or head dimension. Add the architecture-specific cache separately.

Weight formula

Parameter count sets the storage floor.

Raw weight memory is parameters multiplied by bits per weight. Format overhead accounts for scales, metadata and higher-precision tensors; the exact artifact can still differ.

GiB = parameters × bits ÷ 8 ÷ 2³⁰
Why KV is excluded

Parameters cannot predict serving cache.

Two models with the same parameter count can have different layer counts, KV heads and head dimensions. Supplying a fake cache estimate would make the result look more complete and less trustworthy.

Use an architecture-aware cache estimate →
Next step

Replace the generic estimate when the model is known.

Select a verified model in the full worksheet to add context, concurrency, cache precision and multi-GPU sharding.

Open the full VRAM calculator →