---
license: cc-by-4.0
language:
  - en
pretty_name: DGX Spark LLM Arena benchmarks
tags:
  - llm
  - inference
  - benchmark
  - on-prem
  - dgx-spark
  - vllm
size_categories:
  - n<1K
configs:
  - config_name: models
    data_files: models.csv
  - config_name: results
    data_files: results.csv
---

# DGX Spark LLM Arena benchmarks

Reproducible LLM inference benchmarks on an NVIDIA DGX Spark (GB10, 128 GB unified memory). Nine tests per model: five closed-loop (llama-benchy) and four open-loop (vllm bench serve). Quality scores come from vendor model cards.

- **Hardware:** NVIDIA DGX Spark (GB10, 128 GB unified memory)
- **Creator:** Django de Vreng (https://djangodevreng.nl)
- **Visualised:** https://djangodevreng.nl/arena/
- **Raw runs:** https://github.com/djangodevreng/dgx-spark-benchmarks
- **Last updated:** 2026-06-23

## Configs

- `models`: one row per model (17 rows). Metadata, quality scores (MMLU-Pro, GPQA-Diamond, HumanEval/LiveCodeBench from vendor model cards) and the leaderboard composite scores per use-case preset.
- `results`: long format, one row per measured (model, benchmark) pair. 9 benchmarks: five closed-loop (llama-benchy) and four open-loop (vllm bench serve).

```python
from datasets import load_dataset

models = load_dataset("Djangodevreng/dgx-spark-benchmarks", "models")
results = load_dataset("Djangodevreng/dgx-spark-benchmarks", "results")
```

## Method

Three runs per benchmark, fixed seed (42), reported as the mean. Run-to-run variance stays within about 2%. No latency gate: slow models stay visible. Full methodology and the raw stdout per run live in this repo.

## License

[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). Free to use, including commercially, with attribution: Django de Vreng, https://djangodevreng.nl.

## Citation

```bibtex
@misc{devreng-dgx-spark-benchmarks-2026,
  author = {Django de Vreng},
  title  = {DGX Spark LLM Arena benchmarks},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Djangodevreng/dgx-spark-benchmarks}
}
```