System
DGX Spark
GB10 Grace Blackwell Superchip
Memory
128 GB
Unified LPDDR5x
Bandwidth
273 GB/s
Primary decode bottleneck for dense models
Best measured
52.7 tok/s
Qwen3.6-35B-A3B FP8, vLLM
Decision Table
Click column headers to sort. Click a row to inspect evidence, trade-offs, and reproducible setup commands.
| Rank | Model | Type | Params | Measured DGX Spark Speed | Quality Score | Engine / Quant | DGX Sparks | Recommendation |
|---|
Selected Model
Evidence Ledger
Inference Engine Recommendations
Engine guidance is summarized from the active DGX Spark benchmark and recipe dataset.
Setup Commands
Copy the baseline command for the selected deployment path. Treat these as starting points and re-run benchmark prompts locally.
Ranked Recommendation
The ranking prioritizes measured DGX Spark speed, setup practicality, and coding/agentic quality.