Why Device Update Tools Don't Manage Model Releases
Deplero Team
Deplero
Most teams shipping AI to physical sites already have an over-the-air update tool. It keeps the operating system patched, pushes containers, and shows which devices are online. So when the question "how do we roll out the new model?" comes up, the natural answer is "the same way we push everything else". It works for the first few releases. Then it starts to hurt, because a model release isn't a file push.
What device update tools do well
Fleet and over-the-air tools are built to keep devices healthy: inventory, OS and firmware updates, atomic image swaps with rollback, remote access, and hardware health. That job matters, and the good tools do it reliably. Keep using one. The problem is asking it to do a second job it wasn't designed for.
What a model release needs that a file push doesn't
- Model-level versions. A device runs several models. You need to know which version of each one runs where, not just which image the device booted.
- Compatibility with the device. The same model needs a different runtime build for different GPUs, and may not fit in the memory of the smallest box. The release has to be matched to each device before it ships.
- Evaluation on the site's data. A container that starts is a successful update. A model that starts can still give worse answers at one site. That needs a shadow test against the current model, not a liveness check.
- Rollout by site, not by device. Customers think in sites: a hospital, a plant, a lab. Waves, holds and approvals usually follow sites and customers, with a canary device first.
- Rollback of the whole bundle. Model weights, runtime engine, pre- and post-processing and configuration have to move together, or a rollback creates a new mismatch.
- Data rules per site. Some customers allow de-identified clips to reach your cloud; others allow only metrics. Where results and evaluation data may go is part of the release, not an afterthought.
How the two fit together
The cleanest split is by responsibility. The device update tool owns the device: OS, drivers, base image, health. A model release layer owns the models: versions, signed bundles, capacity checks, shadow tests, staged rollout and rollback, across the edge, on-prem GPU clusters and cloud accounts. Each does one job well, and neither has to pretend to be the other.
Questions to ask about your current setup
- Can our tooling tell us which version of each model runs on each device?
- What stops a release from reaching a device that can't run it?
- How do we know a new model gives good answers at a site before it goes live there?
- When we roll back, does the configuration roll back with the model?
- Who decides where each site's data may go, and is that enforced or remembered?
If those questions are answered by people rather than by tooling, the release process is living in someone's head. That works for a few sites. It stops working the week that person is on vacation and a customer needs a fix.
INFRASTRUCTURE FOR PHYSICAL AI
Building this is the hard part. That's our part.
Deplero is the platform for running your AI at every site: vision and custom models at the edge, on-prem GPU clusters, and your cloud accounts, in one system. Installed in your environment, run by your team.
See the platform