SparkBench

Benchmarks and fine-tuning studies on the NVIDIA DGX Spark (GB10, 128 GB unified memory). Not that Spark. Reproducible inference benchmarks across models and quantization levels, plus a pre-registered study: can a LoRA fine-tuned 7–14B open model replace the frontier LLM in AgentReview's code-review agent at near-zero marginal cost?

GitHub

Official source

Start locally with the CLI guide