openclaw - 💡(How to fix) Fix [Feature]: Community hardware reference: DGX Spark + vLLM local setup [1 participants]

Summary

Link to a tested, open-source DGX Spark + vLLM + OpenClaw reference setup in the local hardware docs.

Problem to solve

DGX Spark (128GB Blackwell, ~$4K-$5K) is shipping to consumers and there's no documented reference for running OpenClaw on this hardware with vLLM. Users buying a Spark for private local AI have to figure out model selection, vLLM flags, memory allocation, and OpenClaw provider config from scratch.

Proposed solution

Link to the tested, open-source reference setup in the local hardware or community references section of the docs:

https://github.com/thatwonguy/spark-sovereign (Apache 2.0)

This repo provides a complete, config-driven setup that takes a DGX Spark from box-open to a fully working OpenClaw agentic stack in under an hour:

• 4 bash scripts → vLLM serving on port 8000 → openclaw onboard → done
• Config-driven model registry (config/models.yml) — swap models by editing one YAML file, no script changes needed
• 3 model releases benchmarked with real tok/s on Spark hardware
• Current: Qwen3.5-35B-A3B-FP8 at ~49 tok/s, 131K context, 111/128GB VRAM
• Confirmed working: tool calling (qwen3_coder), parallel agents, MCP, Telegram, voice, memory, multimodal
• Auto-start on boot via systemd

This gives professionals a fully private local alternative to Claude and ChatGPT API dependencies — no rate limits, no usage costs, no data leaving the machine. OpenClaw is the agentic layer that makes it all work.

Happy to contribute a docs PR if the team prefers that path.

Alternatives considered

1. npm plugin package — overkill for what is infrastructure config, not a library. The setup is bash scripts + YAML, not JS. None-the-less able to pursue this if needed.
2. OpenClaw provider preset JSON — possible but limited. The value is the full end-to-end setup (hardware prep, model download, vLLM tuning, OpenClaw onboard), not just the provider config.
3. Just linking the repo in a docs page — simplest, highest value, lowest maintenance burden for the OpenClaw team. This is what I'm proposing.

Impact

Affected users: Professionals buying DGX Spark hardware who want a fully private local AI setup to replace cloud API dependencies (Claude, ChatGPT). This includes engineers, developers, and teams in regulated industries where data cannot leave the network. or anyone sick and tired of bans, walking on egg-shells around ToS, rate-limiting, privacy concerns, costs, mass-surveillance, etc. People who value freedom, privacy, control, and sovereignty, and pushing open-source forward and making intelligence available for all.

Severity: Blocks workflow — without a tested reference, users spend days or weeks debugging model selection, vLLM flags, SM12.1 kernel issues, and OpenClaw provider config on Spark hardware. I went through continuous model releases (and on-going) and dozens of failed configurations to arrive at a working setups.

Frequency: Every new DGX Spark owner hits this. NVIDIA shipped the hardware in March 2026 and community forums are full of users struggling with the same issues I solved.

Consequence: Without a reference, users either give up on local AI and go back to cloud APIs, or waste significant time on trial and error. A docs link costs the OpenClaw team nothing and saves every Spark owner the same debugging journey.

Evidence/examples

No response

Additional information

No response