DGX Spark ·

DGX Spark — Pareto frontier

Projected model map for the NVIDIA DGX Spark: agent capability vs projected end-to-end task time. Task time is projected from the best measured DGX Spark decode speed (tok/s) and the token counts from Tau2, GAIA, and GDPVal eval runs. Capability is the mean pass rate across those three evals. One point per (model × quantization) package. The Pareto frontier sweeps fastest → slowest, keeping only configs with strictly increasing capability.

model entries projected points on the frontier
Pareto frontier dominated candidate
Model Quant Params Score Task time Decode speed Tokens/task Evals Memory Best eval

Pareto-optimal — no other config is both faster and more capable. Candidate — a projected model point but dominated on the frontier.

Method: projectedDgxSparkPoints projects task time as TTFT + (generated tokens / decode tok/s), weighted by task count across the best Tau2/GAIA/GDPVal runs. markParetoFrontier sorts by time ascending and keeps strictly increasing score.