DGX Spark Evals
A personal collection of benchmark results for small language models and quantizations that fit on a single NVIDIA DGX Spark machine, comparing coding and evaluation scores across several models.
This repository records evaluation runs for models and quantized variants sized to run on one DGX Spark, generated with a benchmark script inside a virtual machine sized at 12 CPUs, 50 GB of disk and 32 GB of memory. Current results cover an evolution-style coding evaluation and an aider polyglot coding benchmark, each measured under two different coding-agent backends, alongside an SWE-oriented benchmark.
An earlier leaderboard, kept for reference, used a separate set of evaluation harnesses covering function-calling, code generation and instruction-following, and lists per-model run times and input and output token counts. The repository is a raw results log rather than a packaged tool, intended to compare which small models and quantization levels perform best on this specific hardware.