Hardware check

Use the right answer for the question you actually have.

VRAM Check separates measured performance from planning guidance. Benchmark the machine you own, or explore a future GPU and model combination without presenting estimates as observed fact.

Measured pathReal local inference, captured by the CLI

Use this for claims about speed, latency, stability, and public rank.

Planning pathCatalog evidence, clearly labeled as guidance

Use this to narrow hardware and model options before you spend.

Choose a path

Three useful questions. Three honest starting points.

The fastest route is not always the benchmark. It depends on whether the hardware already exists.

Measure this machine

Run the shared benchmark locally when you need real throughput, timing, repeatability, and a report tied to the hardware in front of you.

Download the CLI
Plan a GPU purchase

Use the GPU catalog when the question is memory capacity, likely model fit, measured field signals, or which upgrade materially changes the ceiling.

Browse GPU profiles
Shortlist a model

Use the model catalog when you already know your available memory and need to narrow the field by parameter count, quantization, context, or task.

Browse model profiles

Why no instant score?

A manual estimator should not look more certain than its inputs.

Throughput needs a real runtime

GPU name and VRAM alone cannot reliably reproduce drivers, backend behavior, power state, memory pressure, or system contention.

Fit can still be useful

Model size, quantization, and memory architecture are valuable planning signals as long as the site keeps them separate from measured speed.

The report preserves the distinction

Every public result keeps its trust state visible so a compatibility estimate cannot silently become a ranked benchmark claim.