Gemma 4 VLA Demo on Jetson Orin Nano Super

Source: HuggingFace Date: 2026-04-22 URL: https://huggingface.co/blog/nvidia/gemma4

Summary

Integration tutorial: NVIDIA’s Asier Arranz builds a full local voice-vision agent on Jetson Orin Nano Super (8GB) using Gemma 4 E2B (Q4_K_M, 5B) via llama.cpp. Pipeline: Parakeet STT → Gemma 4 → optional webcam via look_and_answer tool → Kokoro TTS. Key design: Gemma autonomously decides when to invoke the webcam tool based on conversational context. All components local, no cloud dependency.

Implications

Thread: open-weights ecosystem health / agentic patterns. The autonomous tool-routing design — exposing a single look_and_answer tool and letting Gemma decide when vision is needed — is an elegant agent architecture for resource-constrained hardware: the model only incurs vision inference cost when the task requires it. Gemma 4 at Q4_K_M fitting within 8GB Jetson Orin with multimodal capability demonstrates that the edge VLA hardware requirement is dropping to consumer-accessible levels. The native llama.cpp build requirement (for mmproj support) is a practical friction point that Docker images haven’t caught up to.