---
title: Host your own AI model
description: How to host your own AI model on hardware you control: the model, the system around it, the people it needs, and how to own the stack without running it yourself.
canonical: https://privatesuperintelligence.si/host-your-own-model/
last-updated: 2026-11-20
---

# Host your own AI model

To host your own AI model, you run an open-weight model on hardware you control, in your office, a data center or your own cloud account, so your prompts and documents never pass through a public AI provider. The model is the easy part. Making it useful takes serving, retrieval over your own documents, permissions, security and someone on call. You can staff that yourself, or own the stack and have it run for you.

## What hosting your own model means

- **Your hardware.** The model runs on servers you own or rent, on-premises or in your own cloud account.
- **Your weights.** You download an open-weight model and keep it. OpenAI's gpt-oss-120b is released under the Apache 2.0 license and, per its model card, runs on a single 80 GB GPU. ([Hugging Face](https://huggingface.co/openai/gpt-oss-120b))
- **Your rules.** No vendor sees the traffic, and no vendor can switch it off.

## What it takes

| Part | What it involves |
|---|---|
| Hardware | GPUs sized for the model and the number of people using it |
| Serving | Software that loads the model and handles requests |
| Memory and retrieval | Indexing your email and documents so answers cite them |
| Permissions | Who can see what, and what it may do without asking |
| Security | Patching, monitoring and defenses against prompt injection |
| Upkeep | Testing and swapping in better models as they ship |

A model on your own hardware keeps the provider out. It does not stop a malicious email from instructing the model to forward a file: OWASP lists prompt injection as the first risk for LLM applications. ([OWASP](https://genai.owasp.org/llm-top-10/))

## Hosted on your hardware vs. the frontier labs

Even "self-hosted" offers from the frontier labs keep the model outside your walls. Anthropic's self-hosted Managed Agents run the agent in your network, but "the model runs at Anthropic or at a provider that you choose." ([Claude docs](https://platform.claude.com/docs/en/managed-agents/self-hosted)) When you host your own model, the model call never leaves.

Today's open models that fit on one office server do not match the best closed models on general benchmarks. What closes the gap for your work is context: a model that knows your documents, deals and people. That is what Cortex, the memory layer under Private SuperIntelligence, is built for. In an independent evaluation of agent memory systems, Cortex answered 96% of tasks correctly with no confidently wrong answers. ([Mitosis Labs research](https://mitosislabs.ai/research))

## Own it, without running it

Private SuperIntelligence sets up the model, memory, permissions and interface on infrastructure built for you, on your hardware or in an isolated environment, and runs it with a named contact at Mitosis Labs. You keep the model, the data and the decisions. Delivery takes weeks. For the full staffing comparison, read [Private LLM vs. a private AI service](/blog/private-llm-vs-private-ai-service/) and [Building private AI in-house vs. a done-for-you service](/blog/build-vs-buy-private-ai/).

## Frequently asked questions

### Can I host my own AI model?

Yes. Open-weight models such as gpt-oss-120b can be downloaded and run on a single 80 GB GPU. Running it for daily use also takes serving software, retrieval, permissions and security.

### Is a self-hosted model as good as ChatGPT?

Not on general benchmarks today. On questions about your own documents and decisions, context matters more than raw model size, which is why memory such as Cortex matters.

### What hardware do I need to host my own model?

It depends on the model and the number of users. A large open model such as gpt-oss-120b fits on one 80 GB GPU; the smaller gpt-oss-20b fits within 16 GB of memory.

### Does hosting my own model stop prompt injection?

No. It keeps the provider out, but a document or email can still carry instructions to the model, so permissions and approvals still matter.

### Can Private SuperIntelligence run on my hardware?

Yes. Each deployment is built for one client, from the hardware to where the data is stored, including air-gapped options.
