NVIDIA brings agents to life with DGX Spark and Reachy Mini

Source: HuggingFace Date: 2026-01-05 URL: https://huggingface.co/blog/nvidia-reachy-mini

Summary

Integration tutorial: demonstrated at CES 2026, building a physical AI agent on Reachy Mini + DGX Spark using NVIDIA NeMo Agent Toolkit. Stack: Nemotron 3 Nano 30B (reasoning), Nemotron Nano 2 VL 12B (vision), Pipecat (real-time voice/video), and Reachy Mini (physical action). Smart routing dispatches text queries, visual questions, and tool-requiring tasks to appropriate models. Local footprint: ~65GB + ~28GB.

Implications

Thread: agentic patterns / open-weights ecosystem health. The smart routing architecture is the interesting pattern: instead of one large multimodal model handling everything, specialized models are composed with a routing layer. This is a practical engineering decision — Nemotron 3 Nano 30B for reasoning and a separate 12B VLM for vision is cheaper and faster than a single large VLM at equivalent quality. The Pipecat real-time voice layer solving low-latency multimodal orchestration is a useful open-source component. The CES 2026 debut is primarily symbolic — NVIDIA and HF demonstrating the open-stack personal AI narrative against closed assistants.