HomeAIRoboticsNVIDIA Jetson Demo Runs a Gemma 4 Vision Kiosk On-Device

NVIDIA Jetson Demo Runs a Gemma 4 Vision Kiosk On-Device

Updated 8 November 2026: Added analysis and related TSN links. NVIDIA’s Jetson AI Lab tutorial for the Codex kiosk demo remains published; it is still a vendor walkthrough, not a product launch.

At 19:59 BST on 8 October 2026, NVIDIA Robotics posted on X: “Use Codex to deploy a Gemma4-powered vision demo as a standalone kiosk on NVIDIA Jetson. Automatic startup, live webcam descriptions, and local inference. No Mac required” [1].

This is a vendor tutorial and demo, not a product launch. It shows how to put existing tools together. No new hardware or model was announced.

What the Jetson tutorial shows

The post links to a six-minute NVIDIA Developer video, “How to Deploy a Standalone VLM Kiosk on NVIDIA Jetson with Codex”, published on 7 October [2]. TSN has not watched the video. This account relies on its official description and on the written walkthrough on NVIDIA’s Jetson AI Lab site [2][3]. As of 8 November 2026 that written guide is still live and still presents the Codex session as the recorded example, while listing other coding assistants as planned walkthroughs rather than documented alternatives.

According to those sources, the parts are:

  • The device: an NVIDIA Jetson, the company’s small computer for robots and edge devices. The recorded setup uses a Jetson Orin Nano Developer Kit [3][4].
  • The model: Gemma 4 E2B, the smallest model in Google’s open Gemma 4 family. Google describes it as having “2.3B effective” parameters (5.1 billion including embeddings). It takes text, image and audio input and produces text [5]. Google says its smaller Gemma 4 models are “specifically designed for efficient local execution on laptops and mobile devices” [5]. Jetson AI Lab positions E2B as its best fit for Orin Nano [9].
  • The app: Live VLM WebUI, an open-source NVIDIA tool that streams a webcam to a vision language model (VLM) and shows the model’s descriptions live [2][6]. A VLM is a model that can look at an image and write about it.
  • The agent: OpenAI’s Codex, “a coding agent from OpenAI that runs locally on your computer” [7]. NVIDIA’s prompts ask it to plan the demo from prebuilt containers, deploy Live VLM WebUI, check live webcam captions, and set the Jetson to start the full-screen demo on boot. A reboot test confirms fresh captions with no clicks, and with the host Mac disconnected [2][3].

What “No Mac required” means. It refers to the finished kiosk, not the setup. NVIDIA’s recording uses a Mac to drive the Codex session [3]. Once set up, the Jetson should run the demo alone, “without this Mac or internet” [3]. The agent is a setup tool, not part of the running kiosk.

The runtime. NVIDIA’s prompts point Codex at the Jetson AI Lab Gemma 4 E2B page, which lists vLLM and llama.cpp containers [3][8]. Its Gemma 4 guide calls llama.cpp “the straightforward path” on Orin Nano [9]. The written material does not name which engine the recording used.

Why local inference on the device matters

“Inference” means running a trained model to get an answer. Doing it on the device, rather than sending data to a cloud server, has three plain advantages:

  • Privacy. The camera images are processed on the Jetson, so they do not need to leave the building. That matters for shops, schools, clinics and homes.
  • Latency. There is no round trip to a distant data centre, and the device keeps working when the network is slow or down.
  • Cost. There is no per-request cloud bill. The trade-off is buying the hardware and living with a much smaller model than the cloud offers.

A second point: the fiddly setup can be described in plain language and carried out by an agent, with a person approving each step.

The safety notes NVIDIA itself lists

Kiosk mode usually means automatic desktop login. NVIDIA tells readers to use a dedicated demo account, review auto-login and startup services, and keep rollback instructions. It also warns that “prompt wording is not a security boundary”, so users should set the agent’s real permissions and check what it changed. Its five-minute check is “a smoke test, not proof of production reliability” [3].

A Jetson kiosk demo is a wiring diagram, not a product launch

The useful part of this tutorial is how ordinary the ingredients are. NVIDIA did not announce a new board, a new model or a new software product. It showed a coding agent assembling a kiosk from parts that already existed: a Jetson Orin Nano, Google’s small Gemma 4 E2B vision-language model, the open-source Live VLM WebUI, and prebuilt containers. Once the device is set up, captions are generated on the Jetson itself. That is the whole claim, and it is enough without dressing the post up as a launch.

Local inference is the reason the demo is worth watching at all. A shop, clinic or classroom that wants a camera to describe a scene does not always want those frames to leave the building. A device that keeps working when the network drops is also a different product from a cloud demo that dies with the Wi-Fi. The trade-off is explicit in the model choice: Gemma 4 E2B is the small end of Google’s family, picked because it fits on an Orin Nano, not because it matches a large cloud model. TSN has covered the same on-device direction in NVIDIA’s RTX Spark laptops for local AI agents, the compact Nemotron 3.5 Bolt E4B model that NVIDIA says fits in about 4GB, and Google’s own EmbeddingGemma 2 work on phones.

Two limits should stay attached to the story. First, Codex is used during setup; NVIDIA’s “No Mac required” line describes the finished kiosk, not an agent that configures the board without a host computer. Second, NVIDIA’s own page says a working demo still needs product-specific validation, that a five-minute caption check is only a smoke test, and that prompt text is not a security boundary. Automatic login and startup services are convenient for a kiosk and risky on a machine that also holds other projects. Treat the walkthrough as a wiring diagram for local vision, not as evidence that the arrangement is ready to deploy.

Sources

  1. NVIDIA Robotics (@NVIDIARobotics), X post, 8 October 2026, 19:59 BST (read via the X API). https://x.com/NVIDIARobotics/status/2108271171275809191
  2. NVIDIA Developer, “How to Deploy a Standalone VLM Kiosk on NVIDIA Jetson with Codex” (YouTube video; title and description only; video not viewed by TSN), published 7 October 2026. https://www.youtube.com/watch?v=r_VP4E_Y3as
  3. NVIDIA Jetson AI Lab, “A New Way to Build Edge AI” (AI-assisted development on Jetson tutorial, including setup and kiosk prompts). https://www.jetson-ai-lab.com/tutorials/ai-assisted-development-on-jetson/
  4. NVIDIA Developer, “How to Set Up NVIDIA Jetson for Remote AI Development with Codex” (Part 1, YouTube; description only), published 2 October 2026. https://www.youtube.com/watch?v=CLGaG0JeUBI
  5. Google AI for Developers, “Gemma 4 model card” (last updated 30 July 2026). https://ai.google.dev/gemma/docs/core/model_card_4
  6. NVIDIA-AI-IOT, “Live VLM WebUI” (GitHub repository, Apache 2.0). https://github.com/NVIDIA-AI-IOT/live-vlm-webui
  7. OpenAI, “Codex” (GitHub repository README). https://github.com/openai/codex
  8. NVIDIA Jetson AI Lab, “Gemma 4 E2B” (model page). https://www.jetson-ai-lab.com/models/gemma4-e2b/
  9. NVIDIA Jetson AI Lab, “Gemma 4 on Jetson” (tutorial). https://www.jetson-ai-lab.com/tutorials/gemma4-on-jetson/

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. WashU Makes Edible Fungal Protein From Waste CO2
    AI Science & Health

More in this category

Latest stories

  1. WashU Makes Edible Fungal Protein From Waste CO2
    AI Science & Health

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.