TextGen is a desktop application and web UI for running large language models locally. It offers chat and free-form text generation, image attachments for multimodal models, file attachments (text, PDF, and .docx), tool calling, and web search, and it runs offline with no telemetry, external resources, or remote update checks.
Portable builds for Linux, Windows, and macOS ship with CUDA, Vulkan, ROCm, and CPU-only variants and run GGUF models through llama.cpp; you download a model and drop it in the models folder for the UI to detect. A full installation adds the Transformers, ExLlamaV3, ik_llama.cpp, and TensorRT-LLM backends, LoRA training, image generation, and extensions such as TTS, voice input, and translation. Backends and models can be switched without restarting.
An OpenAI- and Anthropic-compatible API exposes Chat, Completions, and Messages endpoints with tool-calling support, so the app can stand in for those APIs locally. Tools are single .py files, and MCP servers are also supported.
Features
- Chat modes: instruct for instruction following and chat/chat-instruct for custom characters, with prompts formatted by Jinja2 templates
- Multimodal input: attach images to messages for models with vision support
- File attachments: discuss the contents of text, PDF, and .docx files
- Conversation editing: edit messages, move between message versions, and branch a conversation at any point
- Notebook tab: free-form generation outside chat turns
- Backends: llama.cpp, ik_llama.cpp, Transformers, ExLlamaV3, and TensorRT-LLM, switchable at runtime
- Compatible API: OpenAI and Anthropic style endpoints with tool calling for local drop-in use
- Tool calling and MCP: models call custom Python tools such as web search, page fetching, and math, or MCP servers
- Training: fine-tune LoRAs on multi-turn chat or raw text datasets with resumable runs
- Image generation: a tab for diffusers models such as Z-Image-Turbo with 4-bit/8-bit quantization and a persistent gallery
- Interface: dark and light themes, code syntax highlighting, LaTeX rendering, and built-in or community extensions
- Privacy: fully offline operation with zero telemetry
