NVIDIA just showed how to build your own desk buddy robot with DGX Spark

Jensen Huang walked onto the CES 2026 stage and didn’t just talk about AI agents — he showed one you could build yourself. The demo paired a DGX Spark, NVIDIA’s desktop AI computer, with a Reachy Mini robot to create what amounts to a personal office R2-D2. It can see you, talk to you, and actually do things. No cloud required.

The recipe is refreshingly straightforward. NVIDIA published the full source code and a step-by-step guide using their NeMo Agent Toolkit to wire everything together. The brain is a Nemotron 3 Nano reasoning model, eyes come from a Nemotron Nano 2 VL vision model, and the voice is ElevenLabs. Everything runs locally if you’ve got the hardware — about 65GB of disk space for the reasoning model and another 28GB for vision. If you don’t have a DGX Spark sitting around, you can deploy on cloud GPUs through NVIDIA Brev or Hugging Face Inference Endpoints, or just hit serverless endpoints at build.nvidia.com.

What makes this different from yet another AI assistant demo is the commitment to openness. Every component is swappable. Don’t like Nemotron? Swap in another model. Want different tools or behaviors? Change the prompts and routing logic. The NeMo Agent Toolkit handles orchestration but stays framework-agnostic — it plays nice with LangChain, CrewAI, whatever you’re already using. There’s even built-in profiling to track token usage and latency, then auto-tune things to cut costs without sacrificing accuracy.

This isn’t a polished consumer product. It’s a builder’s toolkit, and that’s exactly the point. NVIDIA is betting that the next wave of useful AI won’t come from one company shipping a closed assistant, but from thousands of developers wiring together open models on hardware they control. Whether that bet pays off depends on whether the experience is good enough to make people want a robot on their desk — and whether the setup process stays this simple when you try to make it do something genuinely useful.