# dgx-spark-fleet
An agent skill for setting up and running 1 to 4 NVIDIA DGX Sparks: first boot, linking them, the best models for each fleet size, and fixes.
The skill reads only what it needs. `SKILL.md` asks how many Sparks you have, then opens one file:
```
skills/dgx-spark-fleet/
├── SKILL.md # router + key facts
└── references/
├── 1-spark.md # first boot, remote access, first model, one-Spark recipes
├── 2-sparks.md # cable, netplan, SSH, NCCL test, two-Spark recipes
├── 3-sparks.md # three-cable ring, three-Spark recipes
├── 4-sparks.md # switch or switchless ring, four-Spark recipes
├── models.md # model index: best model per fleet size, measured speeds
└── troubleshooting.md # common failures and fixes
```
## Install
Copy the skill folder into your agent's skills directory:
```bash
git clone https://github.com/0xSero/dgx-spark-fleet.git
cp -r dgx-spark-fleet/skills/dgx-spark-fleet ~/.claude/skills/ # Claude Code
# or ~/.codex/skills/, ~/.omp/agent/skills/, or your project's skills folder
```
Then ask your agent things like "set up my two Sparks" or "what's the best model for three Sparks?".
## Best model per fleet size (September 2026)
| Sparks | Model | Single-user speed |
|---|---|---|
| 1 | Qwen3.6-35B-A3B (NVFP4) | 95 tok/s |
| 1 | Qwen3.8-Flash-Next (NVFP4) | 49 tok/s on prose |
| 2 | GLM-5.3-Flash (EXL3 4-bit) | 27 prose, 65 JSON |
| 3 | DeepSeek-V4.1-Flash (native) | 51 prose |
| 4 | DeepSeek-V4.1-Flash (TP4) | 59 prose, 110 code, 121 JSON |
Full index with recipes and sources: [`references/models.md`](skills/dgx-spark-fleet/references/models.md).
## Sources
- NVIDIA [DGX Spark playbooks](https://build.nvidia.com/spark) ([GitHub](https://github.com/NVIDIA/dgx-spark-playbooks)) for every setup and networking command
- [MiaAI Lab](https://github.com/MiaAI-Lab) recipes, [Local Inference Lab](https://github.com/local-inference-lab) kernels (b12x)
- My speed tests: [local-ai-registry](https://github.com/0xSero/local-ai-registry)
- Longer write-up: *An introduction to the DGX Spark* (Notion)
## License
MIT