Skip to content
,

Text Generation WebUI Guide (2026): Install oobabooga TextGen, Models and API

8–11 minutes
Surprised woman in a green blazer holding a TEXTGEN card in a dark blue studio, hook text: Local ChatGPT one click

Text Generation WebUI by oobabooga has been one of the best-known ways to run AI chat models on your own PC for years.

In 2026 it has a new name, TextGen, and a much easier setup: you can now download a portable app, unzip it, double-click and start chatting, with no Python or command line needed.

It is fully open source with zero telemetry, runs GGUF models through llama.cpp out of the box, and adds vision, file attachments, tool calling with MCP, LoRA training, image generation and an API that works as a drop-in replacement for both the OpenAI and Anthropic APIs.

This guide covers which version to install, downloading your first model, the settings that matter, using the API, and fixing common problems. Details come from the official oobabooga/textgen README and wiki, checked in November 2026.

TL;DR

  • New name: text-generation-webui is now TextGen at github.com/oobabooga/textgen.
  • Easiest install: download a portable build (CUDA, Vulkan, ROCm or CPU) from the releases page, unzip, double-click textgen.
  • Models: put a .gguf file in user_data/models; it appears automatically.
  • Full install (Transformers, ExLlamaV3, training, image generation, extensions): run start_windows.bat, pick your GPU, open http://127.0.0.1:7860. Needs about 10 GB of disk.
  • API: add --api; it serves on port 5000 with OpenAI- and Anthropic-compatible endpoints.
  • Flags go in user_data/CMD_FLAGS.txt.

What Is TextGen (Text Generation WebUI)?

TextGen is a local app for large language models made by oobabooga. Its README sums it up as “a desktop app for local LLMs. Open source, no telemetry.” It handles text, vision, tool calling and web search, with both a user interface and an API.

Feature Details
Back ends llama.cpp, ik_llama.cpp, Transformers and ExLlamaV3
Chat modes instruct (ChatGPT-style), chat and chat-instruct for custom characters; templates applied automatically with Jinja2
Multimodal Attach images to messages with vision models
Files Upload text, PDF and .docx files to discuss them
Tools Web search, page fetching, maths and custom Python tools; MCP servers supported
Training Fine-tune LoRAs on chat or raw text datasets
Images Image generation tab for diffusers models such as Z-Image-Turbo
API OpenAI- and Anthropic-compatible Chat, Completions and Messages endpoints
Privacy 100% offline, no telemetry, no external resources or update checks

You can still find it by its old name, and links to github.com/oobabooga/text-generation-webui redirect to the new repo.

Portable vs Full Install

Portable build Full install
Setup Unzip and double-click One-click script downloads PyTorch and dependencies
Disk Small About 10 GB plus models
Model formats GGUF (llama.cpp) GGUF, Transformers (safetensors), EXL3
Training, image generation, extensions No Yes
Interface Desktop window Browser at port 7860
Best for Most people who just want to chat Power users, fine-tuning, EXL3 speed

Install the Portable App

  1. Open github.com/oobabooga/textgen/releases.
  2. Download the build for your system: Windows, Linux or macOS, with CUDA (NVIDIA), Vulkan (most GPUs including AMD and Intel), ROCm (AMD on Linux) or CPU-only.
  3. Unzip it to a folder with space for models, for example D:\AI\textgen.
  4. Double-click textgen. A window opens with the chat interface.
  5. Add a model (next section) and start chatting.

Everything the app needs is included. If you would rather use it in your browser, the README also shows a venv-based portable install: clone the repo, create a virtual environment, install requirements/portable/requirements.txt and run python server.py --portable --api --auto-launch.

Full Install With the One-Click Script

  1. Clone the repo (git clone https://github.com/oobabooga/textgen) or download the source ZIP and extract it.
  2. Run the start script for your OS: start_windows.bat, start_linux.sh or start_macos.sh. No admin rights needed.
  3. When asked, choose your GPU vendor.
  4. Wait while it sets up a Conda environment (via Miniforge) in installer_files and downloads PyTorch.
  5. Open http://127.0.0.1:7860 in your browser.

To start it again later, run the same start script. To open a shell inside its Python environment, use cmd_windows.bat (or the Linux/macOS version). A Docker Compose setup is also available.

Download and Load a Model

  1. Find a GGUF model on Hugging Face (search “gguf” and sort by downloads). Our guide to downloading models from Hugging Face explains how.
  2. Pick a quantisation that fits your VRAM. Q4_K_M is a good default; our GGUF quantization guide explains the options. The project also links a GGUF memory calculator on Hugging Face.
  3. Put the .gguf file in user_data/models.
  4. Open the Model tab, select it in the dropdown (click refresh if needed) and click Load.
  5. Check the GPU layers and context size before loading. More GPU layers means faster replies; a bigger context uses more memory.

Multi-file models (16-bit Transformers or EXL3) go in their own subfolder inside user_data/models, for example user_data/models/Qwen_Qwen3-8B/ with the config, tokenizer and safetensors files. These need the full install.

VRAM Model size to try (GGUF Q4_K_M)
6-8 GB 3-8B models
12 GB Up to about 14B
16 GB Up to about 20-24B
24 GB+ 27-32B, or larger MoE models with --cpu-moe offloading

Chat Modes and Key Settings

Mode Use it for
instruct ChatGPT-style assistant answers
chat Talking to a custom character with its own persona
chat-instruct Characters, but with the model’s instruction template applied for better quality
Notebook tab Free-form text continuation, outside chat turns
  • Templates are automatic. TextGen reads the chat template from the model metadata, so you rarely need to pick one manually.
  • Sampling: the Parameters tab has presets. For most modern models, use the values the model card recommends (often temperature 0.6-0.8, top_p 0.95, top_k 20 or min_p).
  • Thinking models: there are options to enable or disable thinking and set reasoning effort for models that support it.
  • Branching: edit any message, switch between reply versions or branch the conversation.

Vision, Files, Tools and Image Generation

  • Vision: load a multimodal GGUF together with its mmproj file, then attach images to your messages.
  • File attachments: upload a PDF, .docx or text file and ask questions about it.
  • Tool calling: enable tools such as web search and page fetching. Each tool is a single .py file, so you can write your own, and MCP servers can be added.
  • Image generation (full install): a dedicated tab for diffusers models like Z-Image-Turbo with 4-bit or 8-bit quantisation and a gallery that keeps metadata.
  • Training (full install): train LoRAs on your own chat logs or text, with resume support.
  • Extensions: TTS, voice input, translation and many community extensions.

Use the OpenAI/Anthropic-Compatible API

Add --api to your flags (for example in user_data/CMD_FLAGS.txt) and restart. The API runs at http://127.0.0.1:5000, fully offline, without logging and without connecting to OpenAI.

Test it with:

curl http://127.0.0.1:5000/v1/chat/completions -H "Content-Type: application/json" -d '{"messages":[{"role":"user","content":"Hello!"}],"temperature":0.6}'

  • Full endpoint docs: http://127.0.0.1:5000/docs
  • Change the port: --api-port 1234
  • Require a key: --api-key yourkey
  • Allow other devices on your network: --listen
  • Public Cloudflare URL: --public-api (only if you understand the risk, and always with an API key)

Because it speaks the OpenAI format, apps that let you set a custom base URL (such as SillyTavern, coding assistants or your own scripts) can use http://127.0.0.1:5000/v1. It also exposes an Anthropic-style Messages endpoint, so some Claude-compatible tools can point at it too.

Useful Command-Line Flags

Flag What it does
--api Enable the API on port 5000
--model NAME Load a model at startup
--ctx-size N Context length for llama.cpp models
--gpu-layers N How many layers to put on the GPU
--cpu-moe Keep MoE expert weights in RAM to fit big MoE models
--tensor-split Split a model across several GPUs
--listen / --listen-port Serve the UI on your network or another port
--gradio-auth user:pass Password-protect the web UI
--auto-launch Open the browser automatically
--idle-timeout N Unload the model after N minutes idle

Pass flags directly to the start script (./start_linux.sh --help lists them all) or save them in user_data/CMD_FLAGS.txt.

Update and Reinstall

  • Portable: download the new release and move your user_data folder (models, characters, settings) across.
  • Full install: run update_wizard_windows.bat (or the Linux/macOS version). It also has an option to install or update extension requirements.
  • Fresh reinstall: delete the installer_files folder and run the start script again. Your user_data folder stays.

Troubleshooting

Problem Likely cause Fix
Model not in dropdown Wrong folder Put .gguf files directly in user_data/models, then click refresh
Out of memory when loading Too many GPU layers or big context Lower GPU layers or ctx-size, or use a smaller quant
Very slow replies Running on CPU Use the CUDA/Vulkan build and raise GPU layers
Safetensors model won’t load in portable Portable is GGUF-only Use a GGUF version or the full install
Image input ignored No mmproj file loaded Download and select the model’s mmproj file
API not reachable –api missing or wrong port Add –api, check port 5000 or your –api-port
Other devices can’t connect Listening on localhost only Add –listen and protect it with –gradio-auth or –api-key
Broken after update Dependency mix-up Delete installer_files and rerun the start script

TextGen vs LM Studio vs Ollama

TextGen LM Studio Ollama
Open source Yes App is closed source Yes
Interface Desktop or web UI with many options Polished desktop app CLI plus simple app
Back ends llama.cpp, ExLlamaV3, Transformers llama.cpp, MLX Own engine based on llama.cpp
Training and image gen Yes (full install) No No
Best for Tinkerers who want every setting Easiest model browsing Simple model serving for other apps

Compare with our LM Studio guide and KoboldCpp guide if you want a lighter option.

My Take

The portable builds have turned oobabooga’s tool from “for experts only” into something anyone can run. If you want an open-source, no-telemetry app that lets you change every setting, run EXL3 for speed, attach files and images, and even train LoRAs, TextGen is hard to beat.

Start with the portable CUDA or Vulkan build and one Q4_K_M model. Move to the full install only when you want training, image generation or extensions.

FAQ

Is text-generation-webui still maintained?

Yes. It has been renamed TextGen and is actively developed at github.com/oobabooga/textgen, with portable builds for Windows, Linux and macOS.

Is TextGen free?

Yes. It is open source and free, with no telemetry.

Where do I put models in TextGen?

Put GGUF files in user_data/models. Multi-file Transformers or EXL3 models go in their own subfolder there.

What port does text-generation-webui use?

The web UI uses 7860 and the API uses 5000 by default.

Can TextGen run without a GPU?

Yes, with the CPU-only build, but replies are much slower. Small 3-8B models are usable on modern CPUs.

Does TextGen work with AMD GPUs?

Yes. Use the Vulkan build, or ROCm on Linux.

Can I use TextGen with SillyTavern?

Yes. Start TextGen with –api and connect SillyTavern to http://127.0.0.1:5000.

Can TextGen generate images?

Yes, in the full install, through a tab for diffusers models such as Z-Image-Turbo.

Sources

Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Surprised woman in a green blazer holding a TEXTGEN card in a dark blue studio, hook text: Local ChatGPT one click