Skip to main content

Requirements

Check every item before you start installation.

License​

  • An active paid Hosted EDGE license through Google Cloud Marketplace.
  • Free EDGE and sponsored trials do not qualify. The camera rejects Local activation if your entitlement is not a paid Marketplace license.

Network​

  • The camera and the gateway machine must be on the same private IPv4 LAN.
  • The gateway listens on TCP 9443. Confirm nothing else on that machine uses the port, and that no firewall rule blocks it from the camera's subnet.
  • Hostnames and IPv6 are not supported. You must supply an explicit private IPv4 address.
  • The camera's administrator access must be set to HTTPS-only before you import the connection file. See Connect the camera.

Gateway machine​

  • Apple silicon Mac, or Linux on amd64 or arm64.
  • Enough free memory to hold the model and run inference without competing for resources with other workloads. Use a dedicated machine for predictable latency; a machine under other GPU or CPU load will run slower.

Software​

  • Ollama 0.33.3 or newer. This is the version validated for the documented latency figures. Ollama 0.9.3 was measured at 10-11 seconds warm and is unsupported for production use.

  • Gemma 3 4B as the model:

    ollama pull gemma3:4b

  • The local gateway binary, downloaded by the installer. See Install on Mac or Install on Linux.

Performance you should expect​

  • Warm single-shot detection: about 2.8-2.9 seconds, measured on an Apple M3 Max with 36 GB of memory.
  • Cold start (model not yet loaded): 6-8 seconds.
  • Second-stage verification, if you enable it, adds another model call and increases total time.
  • Run a sample detection after any restart of Ollama or the gateway. The first call after a restart is a cold call.

These figures depend on your hardware. A slower or shared machine will be slower. There is no cloud fallback if local latency is high; the request either completes locally or fails locally.

Next step​

Install on Mac or Install on Linux.

Related: anava.ai/edge/local