DGX Spark LLM benchmark
No DGX Spark LLM benchmarks yet. Be the first to submit a signed, reproducible tok/s run from your own DGX Spark.
Will it run local LLMs?
No signed DGX Spark run yet, but its 128GB of unified memory tells you what fits. At 4-bit it comfortably runs models up to about 133B parameters with room for context, and up to roughly 210B with tighter quantization. Check a specific model with the VRAM-fit tool.
For measured decode speed on the nearest hardware we have benchmarked: RTX 5090 · RTX 4090 · RTX 3090. Run the suite on your DGX Spark to fill this page in.
No DGX Spark benchmarks yet.
Run on YOUR hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Common questions about DGX Spark
Direct Q&A drawn from the runs above: fastest LLM, supported model classes, backend rankings, quantization guidance.