NVIDIA shipped NemoClaw on June 1, 2026, a single-command installer for local AI agents on DGX Spark hardware. The same release adds a cluster assistant inside NVIDIA Sync that joins two to four DGX Spark nodes into one inference pool, and ships a Qwen 3.6-35B model variant that runs 2.6 times faster on Spark through vLLM optimizations.
Try It: One-Command Agent Setup
If you already own a DGX Spark, the new path is short. Run the installer from the NemoClaw setup page, pick one of four agent templates, and the system pulls weights, configures vLLM, and registers the agent behind OpenShell sandbox guardrails. The four templates that ship today are Daily Personal News Digest, Software Development Agent, Deck and Document Reviewer, and Calendar Negotiator. Each one is a working starting point you can fork rather than build from scratch, which closes the gap between "I bought a Spark" and "I have a local agent running on my data."
Why It Matters for Creators
Local agent stacks have been the hardest part of DGX Spark ownership. The hardware was easy to plug in, but turning it into something useful meant assembling Triton or vLLM, model weights, an MCP layer, a sandbox, and an agent loop on your own. NemoClaw collapses that into one shell command and a template, and the new multi-node feature means a four-node cluster reaches 512GB of unified memory, enough to host the kind of long-context coding or document agents that previously needed a cloud subscription. The Spark Qwen 3.6-35B NVFP4 build drops first-token latency low enough for interactive use, which is the threshold that makes a local agent feel real instead of demo-grade.
Key Details
NemoClaw installer. Single shell command, pulls model + vLLM runtime + agent template. Available now from build.nvidia.com/spark.
Multi-node clustering. The cluster assistant in NVIDIA Sync handles two to four Spark nodes. Four nodes give 512GB pooled unified memory.
Qwen 3.6-35B-A3B-NVFP4. Spark-optimized variant, 2.6x throughput improvement over the prior build, runs in 4-bit NVFP4.
Out-of-box setup. Over-the-air updates no longer install by default during initial setup, cutting first-boot time.
Agent templates. Four reference agents ship today, all with OpenShell sandboxing for tool calls.
What to Do Next
Owners can register their Spark and pull the new build at build.nvidia.com/spark. Pick a template that matches an existing manual workflow you already do, point it at your data, and time how long the agent takes versus your current cloud loop. The cluster assistant is most useful if you have access to a second Spark on your local network. The 35B Qwen variant is the right starting model unless you specifically need vision, in which case the previous Spark Cosmos and Nemotron builds remain on the release notes page.
