Local LLM Radar

by Fluxterra EU · data verified —

What can I run on my hardware?

Pick a machine and what you'd do with it. The radar ranks every model that fits by capability and how long a task actually takes.

Capability vs throughput

i

Up and to the right is better. Models that don't fit aren't shown.

Horizontal axis is how many of the selected tasks the machine finishes per hour — prompt reading and answer writing combined, so a machine that reads prompts slowly is penalised even if it writes quickly. Vertical axis is capability after the quantization and context corrections (see Methodology). Change the workload above and the ranking rearranges.

Memory vs context length

i

Where the curve crosses the dashed line is the longest context this model can hold on this machine.

Weights are fixed; the KV cache grows linearly with context. The dashed line is the memory actually available to a model on this device — unified-memory Macs expose about 75% by default, discrete GPUs about 94%. The chart follows the model selected in the table below.

Model size vs the memory you have

i

Each bar is what one model needs: weights, KV cache for your context and runtime reserve. Anything crossing the red line will not fit.

Weights are fixed by the model and the quantization; the KV cache grows with context length, which is why a model that fits at 8K can fail at 128K. The solid red line is the memory actually available to a model — on a Mac that is about 75% of the machine's RAM by default, on a discrete GPU about 94%. The faint dotted line is the total the device has, and the gap between the two is what the operating system keeps. Bars past the edge of the chart are labelled with their real size. Click any bar to open its card.

Ranked by capability

Click a row for the full card. Task time splits into waiting for your prompt to be read and waiting for the answer to be written. Evidence tier: measured derived rough.