What Is Text Generation WebUI
Text Generation WebUI is an open interface for working with language models through a browser: a chat mode, a notebook mode for free-form text, and a generation settings panel without a command line. The project supports several backends: llama.cpp for CPU and light GPU servers, and Transformers and ExLlama for full GPU inference.
The tool is handy when a model needs interactive testing, comparing answers at different temperature settings, and connecting extensions — from voice input to image generation. For a plain API without an interface, vLLM fits better, while WebUI covers the scenario of a person working with a model by hand.
Server Requirements
The minimum configuration depends on the chosen backend and model size.
| Scenario | CPU | RAM | GPU |
|---|---|---|---|
| llama.cpp, 7B model | 4 cores | 16 GB | not required |
| Transformers, 7B model | 4 cores | 16 GB | 8 GB VRAM |
| ExLlama, 13B model | 6 cores | 24 GB | 16 GB VRAM |
It is easier to estimate the exact video memory for a specific model in advance — see the article how much VRAM a model needs.
Installing on a VDS
The project ships as a plain Python repository with a dependency installation script.
sudo apt update
sudo apt install -y git python3.11-venv
git clone https://github.com/oobabooga/text-generation-webui
cd text-generation-webui
python3.11 -m venv venv
source venv/bin/activate
pip install -r requirements.txtInstalling dependencies takes a few minutes: the script sets up PyTorch, support for the chosen backends, and a web server on Gradio.
First Launch and Loading a Model
After installation, start the server and specify a port reachable from outside so you can open the interface from your work computer.
python server.py --listen --listen-port 7860 --apiA model loads through the Model tab: GGUF files go into the models/ directory, after which the model shows up in the dropdown list. The --api flag also opens an OpenAI-compatible endpoint, so WebUI can serve as a backend for third-party applications, not just a browser chat.
Useful Extensions and Modes
WebUI covers more than the chat mode — extensions handle most working scenarios:
- Notebook mode — free-form text with no assistant or user roles, handy for continuation and prompt experiments.
- The superbooga extension adds simple document search right inside the dialog.
- Character cards set the model's personality and reply style through a separate JSON profile.
- The built-in API mode lets you connect WebUI to third-party scripts the same way as OpenAI API — though for production load a separate inference server like llama.cpp directly is more reliable.
What to Do About Common Problems
Most failures are fixed by checking dependency versions or launch parameters:
- The interface does not open from outside — check the
--listenflag and the firewall rule for port 7860. - The model fails to load with a memory error — lower the context length in the model settings or pick a lighter quantization.
- The ExLlama backend cannot find the CUDA library — reinstall the PyTorch version listed in the GPU requirements file.
- Answers cut off mid-sentence — increase the max_new_tokens parameter on the generation tab.
Once configured, WebUI turns into a convenient workbench: you can quickly switch models, compare quantization levels, and share access with colleagues on the same VDS.
Frequently asked questions
Can Text Generation WebUI run with no graphics card at all, using only the CPU?
Yes, the llama.cpp backend was built from the start to run without a GPU: a model in GGUF format loads straight onto the processor, and a modest VDS with enough RAM handles models that are not too heavy. Generation speed is noticeably lower than on a graphics card, but for testing and occasional use it still works fine.
How is this interface different from running a separate API server for production load?
WebUI is built for a person working with a model by hand: chat, notebook mode, comparing answers under different settings. The built-in API mode is handy for development and small scripts, but for stable service under real load a dedicated inference server such as llama.cpp directly or vLLM is more reliable, since it spends no resources rendering a web interface.
How can you figure out in advance how much video memory a given model needs before loading it?
Video memory depends mainly on the model's size and the chosen quantization level: the more compressed a model is, the less VRAM it needs, at some cost to answer quality. Before renting a server for a specific model, it is worth estimating the required memory ahead of time rather than finding out by trial and error after installing WebUI.
Can one Text Generation WebUI server be shared among several colleagues at once?
Technically yes: the interface runs as a regular web server, and the link can be shared with several people working with the same models on one shared VDS. But there is no built-in separation into individual accounts with personal chat history and roles, so for team use with access control a tool with a full multi-user system is more convenient.