How to Use Text Generation WebUI: Run LLMs Privately on Your Business PCs

What if your team had a ChatGPT-style assistant that never sent a byte of company data to the cloud?

That’s the promise of Text Generation WebUI — now rebranded as TextGen — a free, open-source desktop app that runs large language models entirely on your own computers. No API keys, no per-token bills, no customer data leaving the office. For a small business handling contracts, client records, or financials, that combination is hard to beat.

The project, built by developer oobabooga, is one of the most popular open-source local-LLM projects on GitHub, with over 47,000 stars as of mid-2026. The v4.x series renamed it from text-generation-webui to textgen and turned it into a true desktop app: download a portable build, unzip it, double-click textgen, and a window opens. According to the official README, it is “100% offline and private, with zero telemetry, external resources, or remote update requests.”

Why it matters for small businesses

Most teams reach for ChatGPT or Claude first, and for good reason — they’re easy. But three problems keep coming up in small companies:

• Privacy. Pasting client contracts or HR notes into a cloud chatbot means trusting a third party with your most sensitive text. TextGen keeps everything on your own machines.
• Cost. API pricing looks cheap per call until ten employees use it all day. A local model costs nothing per token after setup.
• Connectivity. Once models are downloaded, TextGen works fully offline — handy for travel, spotty connections, or secure environments.

There is a fourth, less obvious advantage: TextGen ships an OpenAI/Anthropic-compatible API. That means software your team already built against those APIs can be pointed at your own PC instead — a drop-in way to cut API spend on internal tools.

What you need

• A Windows, Mac, or Linux PC. The portable builds cover NVIDIA (CUDA), AMD (ROCm), Vulkan, and CPU-only machines, with all dependencies included.
• 16 GB of RAM is a comfortable minimum; a discrete GPU makes larger models dramatically faster.
• Disk space for the app plus model files — individual GGUF models typically run from a few GB up to tens of GB.

The portable build is the friendly route. The full installation (~10 GB, downloads PyTorch) unlocks extra backends, LoRA training, image generation, and extensions like text-to-speech.

Step 1: Install it (about a minute)

1. Open github.com/oobabooga/textgen/releases and download the portable build for your operating system.
2. Unzip the archive and double-click textgen. A desktop window opens — no terminal required.

Prefer the classic route? Download the source, run start_windows.bat, start_linux.sh, or start_macos.sh, pick your GPU vendor when prompted, and open http://127.0.0.1:7860 in your browser.

Step 2: Download your first model

1. Browse Hugging Face for a GGUF model — Llama, Qwen, DeepSeek, and Phi families all have popular GGUF releases.
2. Drop the downloaded file into the user_data/models folder. The UI detects it automatically.

Two official helpers make sizing painless: the LocalBench newsletter recommends well-tested GGUF quants, and the project’s own GGUF VRAM calculator on Hugging Face Spaces estimates how much memory a model needs. One caution that applies everywhere: check each model’s license before using it for commercial work.

Step 3: Start chatting

TextGen offers three chat modes. Instruct behaves like ChatGPT — give an instruction, get a result. Chat and chat-instruct are built for custom characters and personas, with prompts auto-formatted using Jinja2 templates.

Beyond plain chat, the everyday business features are:

• File attachments — upload text files, PDFs, and Word documents and ask questions about their contents. Contract review and policy lookup without uploading to anyone’s cloud.
• Vision — attach images to messages for visual understanding with multimodal models.
• Conversation control — edit messages, flip between answer versions, and branch a conversation at any point to explore alternatives.
• Notebook tab — free-form drafting outside the chat turn structure, useful for long-form writing.

Power features worth knowing

TextGen goes well beyond a chat window, which is why it appeals to teams that outgrow simpler tools:

• Multiple backends. Switch between llama.cpp, ik_llama.cpp, Transformers, ExLlamaV3, and TensorRT-LLM without restarting — handy when one backend runs a particular model better.
• OpenAI/Anthropic-compatible API. Chat, Completions, and Messages endpoints with tool-calling support, so your existing code and automations can call your local model instead of a paid API.
• Tool-calling and MCP. Models can invoke custom Python tools — web search, page fetching, math — and connect to MCP servers.
• LoRA training. Fine-tune models on your own multi-turn chat or raw-text datasets.
• Image generation tab. Run diffusers models such as Z-Image-Turbo with 4-bit/8-bit quantization and a persistent gallery with metadata.
• Extensions. Community add-ons for text-to-speech, voice input, and translation live in the textgen-extensions repository.

TextGen vs. Ollama vs. LM Studio

All three run models locally and privately, so the choice comes down to who will use it:

• Ollama is terminal-first and the lightest option — ideal for developers and for running models as a server. See our Ollama tutorial.
• LM Studio has the most polished point-and-click interface, the best fit for non-technical staff. See our LM Studio tutorial.
• TextGen offers the most control — multiple backends, training, image generation, extensions — at the cost of a slightly steeper setup. It rewards the team member who likes knobs and dials.

If you want the absolute simplest start, try GPT4All first; graduate to TextGen when you need the API or training features.

Limitations to be honest about

• Hardware is the ceiling. Big models need serious RAM or VRAM, and CPU-only generation can feel slow for long outputs.
• Advanced setup takes effort. The portable build is one click; the full install with extra backends and extensions is a real sysadmin afternoon.
• Local models still hallucinate. Keep a human in the loop for anything client-facing or legally sensitive.
• You own maintenance. Updates, backups, and model license compliance are your job — there is no vendor to call.

What to do next

Download the portable build, grab a mid-size GGUF model (an 8B-class quant is a good first test), and try it on a real internal document — a policy PDF or last quarter’s report. If the team likes it, the next step is enabling the API flag and pointing one internal script at it. That single move is often where a small business first feels the savings: private AI, on hardware you already own, with no meter running.

Sources

• Official README — oobabooga/textgen (features, installation, backends, API, privacy claims): https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/README.md
• Project repository and releases: https://github.com/oobabooga/textgen
• DEV Community review of Text Generation WebUI v4.x (stars, v4.7.3 desktop bundling, v4.9 MTP speculative decoding): https://dev.to/jovan_chan_9500711396d4e6/text-generation-webui-review-2026-3k53