What does a model cost to run?

Pick a model and a machine. Memory, decode speed, energy per token and cost per million tokens appear with their assumptions beside them.

Decode speedModelled

–

Energy per tokenModelled

–

Cost per million tokensModelled

–

Memory

Decode speed, tokens per second

Cost per million tokens

Metered on this machine Measured

Model, one run	tok/s	Wall W	J/token
9.7B dense, Q4_K_M	36.7	140	3.81
31.1B mixture, 3B active, Q4_K_M	67.3	171	2.54

Smart-plug wall power averaged over a single run of 6 to 10 seconds, 2026-10-10 23:26 UTC. Includes the whole machine. The record and its limits, and the results file.