What self-hosted actually means here
The phrase gets stretched. Plenty of tools describe themselves as self-hosted when what they mean is that you host a thin front end which then calls someone else's inference API. Your prompt still leaves the building, and the bill still scales with a token counter you cannot see.
Meraqi is not that. The inference runs locally. The model weights ship with the platform, sit on your disk and are served from your own GPU, and nothing is pulled from a public registry or model hub, at install or afterwards. The platform runs with outbound network access blocked entirely, which is something you can verify with a firewall rule rather than a trust exercise.
Two places it can run
Our service, over the internet
Sign up and sign in at meraqi.ai. It is one multi-tenant service that individuals and businesses all share, with nothing to provision. The right answer if you would rather not own GPUs or think about containers.
Your network, whatever that means to you
A cloud direct connect, a government cloud, a private data center, an in-office server room, or a box under someone's desk. If it sits inside your perimeter, it qualifies. At the far end is a deployment with no egress at all.
There is no confusing middle tier. Either the content travels to us or it does not. If you want the second option and have a reasonably current GPU, tell us what you are running and we will say honestly whether it will work before you spend an evening on it.
The stack, stated plainly
Deployment is containerized. The whole runtime is declared in one place, and that declaration is the single source of truth for what gets started. Every container ships in the release itself, so what you install is what runs until you install something else.
- An inference server for the language side, serving the quantized open-weight model that handles prompt interpretation and the design treatment.
- A diffusion runtime for image generation, driven by a versioned workflow catalog so the exact graph that produced an image is recorded rather than implied.
- A video pipeline built on an open-weight video model.
- The editor, which is where the layered output lands and stays editable rather than arriving as a flat render.
Every service binds to loopback by default. Nothing is exposed on a public interface unless you choose to put it there, and no component phones home, for license checks or anything else.
Model choices live in a deployment catalog that comes with the platform, rather than being hidden in the application, so you can see which weights are in play, swap them for others you prefer, and pin a version that does not move until you move it.
What running it yourself gets you
The model does not change underneath you
Public websites update their models when it suits them, and you are not told when the swap happens. Here the weights are a file on your disk, and they change when you replace them. Updates arrive the same way, on your schedule, and they update the software, not your data.
Nothing is uploaded to generate anything
Source material, prompts, brand assets and output all stay in your storage. There is no upload step to audit, no retention policy to read, and no subprocessor list to review, because there is no third party in the path.
Cost stops tracking headcount
On your own hardware the marginal cost of another generation is electricity. Seat-based pricing exists because the vendor carries your compute. Once you carry it, that logic no longer applies to you.
Your history is a record, not a feature tier
Every object in the system carries a time-ordered UUIDv7, and the v1, v2, v3 you see in the interface are labels over those identifiers. Every design keeps its versions and its lineage, back to the source material it was built from and forward to wherever it was exported. That is stored with everything else on your deployment rather than rented back to you as an audit add-on. The full treatment is on the regulated industries page.
Is Meraqi open source?
No, and it is worth saying that plainly rather than letting the word do work it has not earned.
Meraqi is commercial software that you self-host. The models it runs are open weight, the deployment is yours, the data is yours, and nothing reaches us. What you do not get is the platform source under an OSI-approved license or the right to fork it.
If open source is a hard requirement rather than shorthand for control, the honest answer is that Meraqi will not meet it, and we would rather tell you now than after an evaluation. Genuinely open tools exist for both design and image generation, with OSI-approved licenses, the right to fork, and real communities behind them. If the license is what matters most, that is where your time is best spent.
If what you actually want is control over where your content lives, which model runs, and when anything changes, self-hosted commercial software gets you all three without the license.
Questions we get asked
What happens if you do not have a GPU, whether you can swap the models, whether it generates with outbound access blocked, and how self-hosting changes the price. Those answers sit with the rest of them on the FAQ page, along with what is true wherever you run Meraqi.
Read the self-hosting questionsRun it on your own metal
Join the beta waitlist. If you have a specific card, a specific stack, or a specific constraint, tell us and we will say honestly whether it will work.
Join the beta