Skip to content

On-premise and self-hosted

On-premise and self-hosted AI: what it involves

What running Marella on infrastructure you control involves, in your own data centre or cloud account, and when a managed deployment is the better choice.

By Marella AI · Updated 23 Oct 2026

Short answer

On-premise AI runs on servers your organisation operates in its own building or data centre. Self-hosted is broader: you run the software on infrastructure you control, which can also be your own cloud account. Neither term says where the model runs, whether the system can reach the internet, or who supports it, so agree each of those separately.

01

Four terms that get used as one

Buyers and vendors use these words interchangeably. They answer different questions, and a requirement that says only one of them leaves the others open.

  • On-premise: the servers are in your building or data centre.
  • Self-hosted: you run the software, on-premise or in your own cloud account.
  • Local model: the model runs on hardware you control, so prompts are not sent to an outside provider.
  • Air-gapped: no network route to the internet at all, which needs its own design and testing.
02

When on-premise or self-hosted fits

Choose it when your rules require infrastructure you control and you have a team that can run it. Typical reasons are a network policy that limits outbound connections, an approved list of storage and model providers, or a rule about where data may be held.

  • Data must stay on infrastructure you control
  • Outbound connections are restricted
  • Storage and model providers must be pre-approved
  • You have staff to run, patch and monitor it
03

When managed is the better choice

If a vendor's documented data boundary meets your requirements, a managed deployment saves you running upgrades, monitoring, backups and incident response. Many requirements that start as “it must be on-premise” turn out, once the data flows are drawn, to be about control of the data path rather than the location of the servers.

  • The vendor's documented boundary meets your rules
  • You would rather not run servers
  • You have no out-of-hours operations cover
04

What Marella's self-hosted package includes

Marella ships as a Docker-based self-hosted package that runs on infrastructure you choose. The production layout, supported services, model configuration and prerequisites are confirmed for your environment at architecture review.

  • Application and worker services
  • Database and object-storage dependencies
  • Model and provider configuration
  • Sign-in, secrets and network settings
05

Test it before production

Install it in an environment like the one you will run, then test sign-in, retrieval, backup and restore, upgrades and restricted outbound access against a data-flow diagram you have agreed.

  • A list of every image and dependency
  • Inbound and outbound network rules
  • How support would reach it, if at all
  • A rehearsed restore and upgrade
06

Who runs what

Put a name against each operational job, from patching the servers to answering an incident at two in the morning. Put exceptions in writing.

  • Infrastructure and application operation
  • Monitoring and vulnerability fixes
  • Secrets, backups and incidents
  • Support access and upgrades

Caveats

  1. B1Self-hosted does not automatically mean on-premise, fully local or air-gapped.
  2. B2Running the models locally depends on the configuration and can change what the system can do and how it is supported.
  3. B3Controlling the infrastructure means taking on the work of running it.
  4. B4No service level, layout or data location applies until it is agreed in writing.

Frequently asked questions

What is the difference between self-hosted AI and a local model?

Self-hosted means you run the software on infrastructure you control, on-premise or in your own cloud account. A local model means the model itself runs on hardware you control, so prompts are not sent to an outside provider. A self-hosted deployment can still call an outside model provider, so agree each one separately.

What does an on-premise AI deployment involve?

Running the software on servers your organisation operates in its own building or data centre, and taking on upgrades, monitoring, backups and incident response. Agree separately where the model runs, whether the system can reach the internet and who supports it. Marella's self-hosted package is Docker-based, with prerequisites confirmed for your environment at architecture review.

Can Marella run on-premise?

Yes. The self-hosted package runs on infrastructure you choose, including servers in your own data centre. Prerequisites and who is responsible for what are agreed at architecture review.

Can Marella run with local models?

Where the software runs and where the model runs are separate choices. Marella's model providers are configurable, and a fully local configuration is checked at architecture review for the models and workflow you propose.

Is self-hosted the same as on-premise?

No. Self-hosted software can run in your own cloud account or on your own premises. Say which, and who operates it.

Is on-premise AI more secure than cloud AI?

Not by default. On-premise gives you control and the work of running it; a managed or cloud deployment hands that work to a provider whose controls you need to check. The safer option is the one your team can operate well.

Who applies updates?

That depends on the support and operating model. Agree how new versions are delivered, tested, scheduled and rolled back, and who handles vulnerabilities, before production use.

Related pages

Plan an evaluation

Agree who runs what before you install

Send the infrastructure constraints and the questions your architecture has to answer. Keep the first message non-confidential.

Request an architecture reviewExplore the platform