Open WebUI in 2026: Setup, Backends, RAG and the Licence Catch

Open WebUI is the chat interface most people put in front of a local model. It looks like ChatGPT, runs on your own machine, and talks to whatever engine you already have. The single most common misunderstanding about it is worth settling first: Open WebUI does not run the model, and it uses no VRAM of its own. It is a web app. Every gigabyte of memory in a local AI setup is spent by the engine behind it.

This guide covers what it connects to, how to install it, what it adds beyond a chat box, and one licensing condition that catches teams out above fifty users. Verified against the project’s own repository and documentation on 14 September 2026.

Quick answer

• What it is: a self-hosted web interface for chatting with local or remote models. Current release v0.11.3, published 31 August 2026.
• What it costs in memory: effectively nothing on the GPU. Size your hardware for the backend, not for this.
• What it connects to: Ollama natively, plus OpenAI, Anthropic and any OpenAI-compatible endpoint — which includes vLLM and llama.cpp’s server.
• Who it suits: anyone who wants a shared, multi-user chat UI with document search, rather than a single-user desktop app.
• The catch: the licence forbids removing Open WebUI branding above 50 users. Details below.

It is a front end, not an engine

Open WebUI sits in front of a runner and speaks to it over HTTP. The runner loads the weights, holds the KV cache and does the maths on your GPU. That separation is what makes the memory question simple: if you are deciding whether a model fits your machine, the interface is not part of the calculation.

So the planning order is backend first, interface second. Work out which model your card can hold with the VRAM & KV cache calculator, then pick the engine that suits how you intend to use it, then install this on top. Our comparison of local LLM runners covers that middle step, and how much RAM and VRAM you actually need covers the first.

What you can put behind it

Because the OpenAI API shape is the common denominator, anything that speaks it will work, and you can register several backends at once and switch per chat.

Installing it

Two supported routes. Docker is the one the project recommends, because it pins the Python version for you. The same pattern applies if you are assembling a wider self-hosted stack: self-hosting n8n with a local model installs the same way and carries a licence restriction of its own.

docker run -d -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  -e WEBUI_SECRET_KEY=$(openssl rand -hex 32) \
  --name open-webui --restart always \
  ghcr.io/open-webui/open-webui:main

That serves the interface on http://localhost:3000 and picks up a local Ollama automatically at host.docker.internal:11434. The named volume is what keeps your chats, users and settings across upgrades — omit it and you start clean every time you pull a new image.

The pip route:

pip install open-webui
open-webui serve

This one serves on port 8080 rather than 3000, and connects to Ollama at localhost:11434.

Check your Python version before you use pip. The project supports Python 3.11 and 3.12 only; 3.13 is not supported yet, and 3.11 is what they recommend for production. On a current distribution or a fresh Homebrew install you may well have 3.13 as the default, which is exactly the case where Docker saves you an afternoon.

What it adds beyond a chat box

If it were only a prompt window it would not be worth a container. The parts that earn it are the retrieval stack and the multi-user controls. Those two interact: the default vector store is SQLite-backed and is not fork-safe above one worker, which is covered in our vector database comparison.

Document search

Retrieval is built in rather than bolted on: 13 vector databases are supported, including ChromaDB and PGVector, alongside 8 extraction engines such as Tika, Docling, Azure and Mistral OCR. To keep that extraction on your own machine instead, see local OCR with Ollama. Search is hybrid — BM25 combined with vector search, then cross-encoder reranking — and there is an agentic mode where the model decides for itself when to go and search your documents.

Retrieval quality depends far more on your embedding model than on the interface, and that choice is a real one. We cover the current options in best local embedding models for RAG, and our walkthrough of building a private RAG chatbot sets the retrieval stack up end to end.

Multiple users

This is the main reason to choose it over a desktop app. It ships with RBAC — roles, groups and per-resource permissions, so you can expose particular models only to particular people — plus SSO via OIDC or LDAP against any identity provider, and SCIM 2.0 for automated user and group provisioning. That is a genuine small-team deployment story, not a single-seat tool.

It also does web search with inline citations, so the model can pull in current information and show where it came from.

What changed recently

Version 0.11.0 (27 July 2026) was the largest structural change in months: unified settings, a rebuilt sidebar, remappable keyboard shortcuts, an overhauled model editor, support for OpenAI’s image generation and editing endpoints, and per-chat memory of which model and tools you had selected. The 0.11.1–0.11.3 releases through August were fixes on top, including accessibility work and OAuth disconnect behaviour for MCP tool servers.

The licence condition worth reading

Open WebUI is widely described as open source, and GitHub does not report a standard licence for it — the repository resolves to NOASSERTION. The reason is clause 4.

The terms are BSD-3-Clause with one addition: licensees are prohibited from altering, removing, obscuring or replacing any Open WebUI branding — name, logo, or other identifiers — in any deployment or distribution. There are three exemptions:

1. Deployments with 50 or fewer end users, counted as individual people with direct access, in any rolling 30-day period;
2. Specific prior written permission from the copyright holder;
3. A duly executed enterprise licence.

Outside those, removing the branding is a material breach. In practice: if you are running it for yourself, a family, or a team of a few dozen, white-labelling is permitted. If you intend to rebrand it for a company intranet of two hundred people, or ship it inside a product, you need to talk to them first. Note also that materials under earlier licences keep their original terms, as recorded in the project’s LICENSE_HISTORY, and that contributing requires signing a CLA.

This is the sort of condition that does not appear in a summary table, and it is worth applying the same scrutiny to the models you serve through it — our licence checker covers what each model licence permits, and the models directory lists the terms alongside the memory requirements.

A sensible first setup

1. Decide the model first. Check what your card holds with the VRAM calculator. A 24GB card comfortably runs a good 27B model at 4-bit — see running Qwen3.8 27B locally.
2. Install the engine and confirm it answers on its own port before adding anything on top.
3. Run the Docker command above, with the volume and a real secret key.
4. Open port 3000, create the first account — it becomes the administrator — and confirm your models appear in the picker.
5. Only then turn on document search, and pick the embedding model deliberately.

The per-model memory figures for each engine are on our Open WebUI hub, which pairs the interface with what to run behind it.

FAQ

Does Open WebUI need a GPU?

No. The interface runs fine on a CPU-only machine. The GPU is needed by whatever engine serves the model, which can be on the same machine or another one on your network.

How much RAM does the interface itself use?

Modest, and unrelated to model size — it is a Python web application with a database. Plan your memory around the model instead.

Can it use ChatGPT and a local model together?

Yes. Register an OpenAI or Anthropic key alongside your local backend and switch between them per chat.

Is it genuinely offline?

The chat path is, once the image and model are downloaded. Features that reach outward by design — web search, hosted API models — obviously are not.

Can I remove the branding?

Only under one of the three exemptions above. At 50 end users or fewer in a rolling 30-day window, yes.

Sources

• Open WebUI releases — version and dates (v0.11.3, 31 August 2026)
• Open WebUI License — clause 4 branding condition and exemptions
• Quick start — Docker and pip commands, Python version support, ports
• Features — backends, vector databases, extraction engines, RBAC, SSO, SCIM
• v0.11.0 release notes — interface changes

Updated 14 September 2026. Version numbers and licence terms verified against the project’s repository and documentation on that date; check the sources above before relying on them for a deployment decision.