NVIDIA announced updates to DGX Spark at Computex 2026 focused on making local AI agent development faster and more accessible. The NemoClaw open-source blueprint now installs via a single curl command, bundling Ollama, the Qwen3.6-35B model, and the OpenShell secure sandboxed runtime. Four ready-to-use agent templates are provided (news digest, software dev agent, document reviewer, calendar negotiator). Performance improvements deliver up to 2.6x faster inference on Qwen3.6-35B using NVFP4 quantization and vLLM optimizations. For teams needing more compute, the NVIDIA Sync cluster assistant automates multi-node setup for 2–4 DGX Spark units over ConnectX-7 200 Gbps RoCE networking, enabling up to 512 GB unified memory for running large models like 400B-parameter MoE architectures.
