<!-- LU-CLOUD-FACTS:START -->
## Current cloud offer (2026-09-14)
Hosted includes 11 Flash models; Pro and Max include 12. In app sessions they cost no credits, up to a daily limit per account. GLM 5.3 Flash uses credits on Hosted and joins the daily allowance on Pro and Max. API keys always use credits.
Hosted EUR 19/month: 900,000 shared credits, up to 900 uncensored images OR 90 five-second uncensored image-to-video clips. Each maximum spends the entire shared monthly wallet on one model. These are alternatives, not separate allowances. Input-image generation, character training, LoRA routes and additional operations cost extra. Uncensored video uses an input image and five-second clips. Failed jobs and changing catalog prices can affect results.
https://locallyuncensored.com/cloud-facts.json
Uncensored AI chat online: 48 chat models, all of them in every plan and every pack, and no moderation layer from us on text. https://locallyuncensored.com/uncensored-ai-chat-online/
Uncensored AI image generator online: uncensored models from 1,000 credits per image, only after the account switches its own content policy setting. https://locallyuncensored.com/uncensored-ai-image-generator-online/
Uncensored AI video generator online: image-to-video only, clips are five seconds, the input image is not included. https://locallyuncensored.com/uncensored-ai-video-generator-online/
Uncensored AI API: the desktop app serves an OpenAI-compatible API for every local model at http://127.0.0.1:8129/v1 (token required, Settings, Voice & Remote, Local API); the hosted endpoint is https://lu-labs.ai/api/inference/v1, keys start with lu_ and always spend credits. https://locallyuncensored.com/uncensored-ai-api/
Venice AI alternatives: Venice Pro is 18 USD per month with 100 credits (read 2026-09-22); our credit packs start at 15 EUR and need no subscription. https://locallyuncensored.com/vs/venice/
https://lu-labs.ai/product-facts
<!-- LU-CLOUD-FACTS:END -->
# Locally Uncensored, Full Reference

> Free, open-source (AGPL-3.0) plug-and-play local AI studio for Windows, Linux and Mac: chat with 20+ providers, generate images and video on your own machine (through ComfyUI on Windows and Linux), and code with a built-in agent. One installer, no command line, no Docker. No analytics kit, and no account for local use. Cloud only when you switch it on.

Current version: v3.0.7.

This file is the full, citable reference for Locally Uncensored. Facts are stated plainly so they can be quoted directly.

## What Locally Uncensored Is

Locally Uncensored is a plug-and-play local AI studio: a desktop application for running AI models on your own hardware that installs like a normal program (no command line, no Docker, no config files). It combines four things most local-AI tools keep separate: AI chat, a coding agent, image generation, and video generation, all in one interface. It is free and open source under AGPL-3.0. It runs locally by default. On its own the app uses the network for two things: an update check against GitHub Releases and a count of each press of the Cloud switch. Cloud providers are optional and use your own API keys.

- Name: Locally Uncensored (abbreviated LU)
- Category: local AI studio / desktop AI application
- License: AGPL-3.0 (free, open source)
- Platforms: Windows 10/11, Linux (deb, rpm, AppImage) and, since v3.0.6, macOS 14 or newer on Apple Silicon (dmg). Intel Macs are not supported.
- Website: https://locallyuncensored.com
- Source: https://github.com/PurpleDoubleD/locally-uncensored
- Author: PurpleDoubleD
- Tech stack: React 19, TypeScript, Tailwind CSS 4, Vite 8, Tauri v2 (Rust backend)
- Price: free

## Frequently Asked Questions

### What is Locally Uncensored?
A free, open-source desktop app for running AI locally. It combines chat (20+ provider presets), a coding agent with 14 tools, image generation via ComfyUI (FLUX 2, Juggernaut XL, Z-Image, SDXL), and video generation (Wan 2.1, HunyuanVideo, LTX 2.3, FramePack F1) in one interface. AGPL-3.0 licensed.

### Does it support Qwen 3.8, GPT-OSS and GLM-4.7?
Yes. Qwen 3.8 27B is in the Model Manager since v2.6.6, with working vision: the image tower ships as a separate mmproj file that the download writes next to the model. Official, huihui abliterated and OrcaRouter uncensored builds are there, plus a 9B distill for small cards and two Ollama tags. GPT-OSS-120B and GPT-OSS-20B run via Ollama. GLM-4.7 Flash is supported through Ollama. DeepSeek R1, Llama 4, Gemma 4, Mistral Small 3 and Phi 4 are also ready.

### Can I use it as a ChatGPT or Claude alternative?
Yes. Locally Uncensored works as a ChatGPT and Claude alternative that runs on your own hardware. Use Qwen 3.8, GPT-OSS, GLM-4.7, DeepSeek R1, Llama 4 or Gemma 4 instead, or add cloud providers (OpenAI, Anthropic, OpenRouter, Groq) alongside the local stack.

### Is it really free and offline?
Yes. After setup and model download, no internet is needed for the local providers. Local use needs no account and has no usage limit. The app has no analytics kit and sends no crash reports. On its own it checks GitHub for updates and counts presses of the Cloud switch. Cloud providers are optional; the core runs entirely on your hardware.

### How is this different from Open WebUI, LM Studio, Jan or Msty?
Those tools handle text chat. Locally Uncensored adds a coding agent with 14 MCP tools, image generation via ComfyUI, video creation, mobile remote access via LAN or Cloudflare Tunnel, A/B model comparison, local benchmarking, granular permissions, file upload with vision, and thinking mode, all in one app.

### What hardware do I need?
Text chat: 8 GB RAM. Image generation: an NVIDIA card from 6 GB of VRAM with the starter checkpoint, 8 to 10 GB for FLUX. Video generation: from 6 GB for AnimateDiff and image to video with FramePack F1, from 8 GB for the Wan 2.1 starter, 12 GB and up for Wan 2.2 and HunyuanVideo. Windows 10/11 and Linux are supported.

### What does "uncensored" mean?
It refers to abliterated AI models with artificial refusal behaviour removed. The model responds without refusing or adding disclaimers. Combined with fully local execution, conversations stay private on your machine.

### Does remote access leak data?
Remote access is off until you start it, sits behind a 6-digit passcode, and shows when a device is connected. On your own network the chat goes straight from phone to PC. Over the internet Cloudflare relays it: the traffic is encrypted to and from Cloudflare, not end to end, so Cloudflare could read it.

### Can I run it on macOS?
Yes, since version 3.0.6, in beta. The Mac build is for Apple Silicon and needs macOS 14 or newer. Intel Macs are not supported. The .dmg is signed and notarized. Chat runs on local models through the bundled server with Metal, and Create makes images and video on the Mac itself, without ComfyUI. We tested it on an M5 Pro with 48 GB. Tell us how it runs on yours. Voice has been tried by hand in the app on a Mac and works. Computer control is in the Mac app but not fully tested yet on macOS. Remote Access and the automatic update are not confirmed on macOS. On any Mac you can also use the hosted studio in a browser at lu-labs.ai.

### Is there a mobile app?
Yes. The desktop app hosts a full mobile web app over LAN or Cloudflare Tunnel. From a phone you get chat, Coding Agent, 14 agent tools, plugins and personas.

### Can it generate images and video, not just chat?
Yes. Locally Uncensored is the local AI app in its class that bundles image generation (FLUX, Stable Diffusion XL family) and video generation (Wan, LTX, HunyuanVideo) alongside chat, in one installer with one-click model downloads. Chat-focused apps like LM Studio, Jan and Ollama are text only. Details: https://locallyuncensored.com/blog/local-ai-image-and-video-generator.html

### Is there a German guide? / Gibt es einen deutschen Guide?
Ja. Der deutsche Einsteiger-Guide zu unzensierter lokaler KI: https://locallyuncensored.com/blog/unzensierte-ki-lokal.html. Die App-Oberflaeche ist englisch, die Modelle (Qwen, Llama, Gemma) antworten fliessend auf Deutsch.

### Can I run GLM-5.3 or GLM-5.3-Flash locally?
Only on very large machines. The GLM-5.3 flagship has 753.9 billion parameters and its smallest quantised build is 216.7 GB; GLM-5.3-Flash has 320.8 billion parameters with 18 billion active per token and starts at 93.1 GB, plus a 1.13 GB mmproj file if you want it to read images. Those are file sizes before context and cache. The flagship runs in a stock llama.cpp build because its glm-dsa architecture has been in the main branch since GLM-5.2; Flash reports itself as glm5next and was still waiting on pull request 27754 on 2 September 2026. Flash is MIT licensed, the flagship ships under Z.ai's own glm-5.3 licence. Both are available hosted in LU Labs Cloud on every plan. Full guide: https://locallyuncensored.com/blog/run-glm-5-3-locally.html

### Can I run Kimi K3 locally?
No: Kimi K3 is a ~2.8 trillion parameter Mixture-of-Experts model, and now that the weights are public the sizes are measured: the smallest 1-bit GGUF is 466 GB and Q4 is 1.5 TB, which is datacenter hardware. Hosted access exists at independent hosts and in LU Labs Cloud on every plan. Run strong mid-size open models (Qwen, GLM, Gemma, Ling 3.0 Flash) locally in Locally Uncensored instead. Full guide: https://locallyuncensored.com/blog/can-you-run-kimi-k3-locally.html

### Can I run Ling 3.0 Flash locally?
Yes, on the right machine. Ling 3.0 Flash is a 124B Mixture-of-Experts model with only 5.1B active parameters, MIT licensed; GGUF quants run from about 42 GB at 2-bit to a 78 GB Q4_K_M sweet spot, so 64 GB machines run the small quants and 96 GB machines run Q4 comfortably. A 24 GB GPU plus 64 GB system RAM also works via MoE offload, and generation stays fast because so few parameters are active. Full guide: https://locallyuncensored.com/blog/how-to-run-ling-3-0-flash-locally.html

### Can I run Qwen 3.8 locally?
Yes, the 27B. Qwen 3.8 27B came out on 13 August 2026 under Apache 2.0 and GGUFs appeared the same day: Q4_K_M measures 17.1 GB and wants a 24 GB card, IQ4_XS at 15.7 GB is the largest quant that stays whole on 16 GB, and the 2-bit quants sit at 9.0 to 10.7 GB for 12 GB cards. Add the 0.9 GB mmproj file if you want it to read images. Not the Max: at ~2.4 trillion parameters even 4-bit weights are around 1.2 TB, which is datacenter hardware, and its weights carry a custom qwen3.8-max licence rather than Apache. Setup guide: https://locallyuncensored.com/blog/how-to-run-qwen-3-8-27b-locally.html

### Can I run Qwen-Image 2.1 locally?
Yes, on Windows or Linux with Locally Uncensored 3.0.1 and ComfyUI 0.37.0 or newer. Qwen-Image 2.1 came out on 20 September 2026 and is the first in its series where one set of weights both generates a picture from a prompt and edits a reference image, with no separate edit model. The Model Manager bundle Qwen-Image 2.1 (Generate and Edit) pulls three int8 files totalling 16.1 GB into the right ComfyUI folders: the 7B image model at 7.26 GB as listed, the Qwen3-VL-8B text encoder at 9.35 GB, and the VAE at 676 MB. As a guide to memory rather than a measurement of ours, ai.rs puts the int8 pipeline at roughly 16 GiB in total, so a 24 GB card holds all three parts resident and a 16 GB card usually keeps the text encoder in system RAM; the full bf16 weights want 33 to 40 GB. Defaults follow the official template: 25 steps, CFG 1, euler, simple, 1024x1024. Editing takes one reference image plus a prompt and no mask. The weights carry the Qwen Research License, which covers research and evaluation and does not grant commercial use without a separate license; Qwen-Image 2512 and Qwen-Image-Edit 2511 remain Apache 2.0. Setup guide: https://locallyuncensored.com/blog/how-to-run-qwen-image-2-1-locally.html. What changed in the model: https://locallyuncensored.com/blog/qwen-image-2-1-explained.html

### Can I run DeepSeek V4 Flash locally?
Yes, on the right machine. V4 Flash is a 284B MoE with only 13B active parameters; the 4-bit GGUF is about 155 GB, so a 192 GB Mac Studio or a workstation with 192-256 GB RAM plus a GPU runs it at interactive speeds. Smaller machines can run the 5.3 GB Qwen 3.5 9B V4 distill from the in-app catalog, and the huihui abliterated build (154 GB, 600K+ downloads) is the most popular uncensored model of 2026. Full guide: https://locallyuncensored.com/blog/how-to-run-deepseek-v4-flash-locally.html. V4 Pro (1.6T) is cloud territory: https://locallyuncensored.com/blog/how-to-run-deepseek-v4-pro.html

### Which AI backends does it detect?
It auto-detects 12 local backends: Ollama, LM Studio, vLLM, KoboldCpp, llama.cpp, LocalAI, Jan, TabbyAPI, GPT4All, Aphrodite, SGLang and text-generation-webui. ComfyUI is used for image and video generation. If none are running, the setup wizard offers one-click install links.

### How long does setup take?
About five minutes. There is no Docker, terminal or config file. Run the installer, let the wizard scan for backends, pick a model, and start.

### Is there an uncensored AI image generator online?
Yes, in the LU Labs Cloud, and it is paid, not free. Nine image models render uncensored content: Neta Lumina Uncensored, Z-Image Base and Nucleus at 1,000 credits per image, Krea 2 Turbo at 1,200, Chroma Uncensored and Prefect Pony XL Uncensored at 1,500, JibMix Qwen and MOSS Image at 2,000 and Wan 2.2 Realism at 2,500, all drawn from the same wallet as chat and video. Uncensored output only happens once the account changes its own content policy setting, content involving minors is blocked everywhere with no exception, and photos of real people need their consent. Details: https://locallyuncensored.com/uncensored-ai-image-generator-online/

### Can I generate uncensored AI video online?
Yes, on seventeen models, with one limit worth knowing before you pay: hosted video is image-to-video only, there is no text-to-video, and every clip is five seconds long. You supply the starting image, and it is not included in the price of the clip. LTX 2.3 Uncensored and OpenVideo cost 10,000 credits per clip and Wan 2.2 Uncensored 15,000, up to Seedance 2.5 Uncensored at 90,000. Wan 3.0 Uncensored and Wan 3.0 Prime Uncensored also accept an end image. Details: https://locallyuncensored.com/uncensored-ai-video-generator-online/

### Is there an uncensored AI API?
Yes. The endpoint is OpenAI-compatible Chat Completions at the base URL https://lu-labs.ai/api/inference/v1. You create a key in the "Cloud API keys" section of an lu-labs.ai account, it begins with lu_ and is shown once, and up to five keys can be active at a time. An API key always spends plan credits, never the daily Flash allowance, and it can neither read nor change the account. Details: https://locallyuncensored.com/uncensored-ai-api/

## Features

- Plug-and-play setup wizard that auto-detects 12 local backends
- 20+ provider presets (local and cloud)
- One-click in-app installers for Ollama and ComfyUI
- 75+ one-click model downloads with hardware-aware recommendations
- VRAM tier filter: Lightweight (≤10 GB), Mid-Range (10-16 GB), High-End (>16 GB) across text, image and video models
- Coding Agent with live token streaming, apply-patch, file tree
- Claude Code CLI integration
- Agent Mode with 14 tools, parallel execution, MCP integration, sub-agent delegation, budget caps
- Image generation via ComfyUI (FLUX 2 Klein, FLUX.1, Juggernaut XL, Z-Image Turbo, ERNIE-Image, SDXL, SD 3.5)
- Video generation (Wan 2.1, Wan 2.2, HunyuanVideo 1.5, LTX 2.3, FramePack F1, AnimateDiff, Mochi, Cosmos)
- Image-to-image with denoise control, in a tab labelled "Edit / Image to Image"; image-to-video (FramePack F1 on 6 GB VRAM)
- Local Create lanes for Talking Character, Music, Extend Video and Motion Control, built from core ComfyUI node families and run on your own GPU; Upscale and Erase Object are cloud-only, and Character Studio trains on your own GPU or in the cloud
- Per-lane frame count, resolution and step controls on every Create lane
- Surgical file_edit tool in the Coding Agent that changes the requested lines instead of rewriting whole files
- Mobile remote access via LAN or Cloudflare Tunnel with 6-digit passcode and QR setup, working against non-Ollama backends too (LM Studio, Lemonade, llama.cpp), with streaming, tool calls, vision and reasoning
- A/B model compare and local benchmark
- Memory system and Document Chat (RAG) with local embeddings (nomic-embed-text), no cloud
- Voice (since v3.0.6, beta): a spoken conversation with the selected chat model. Speech recognition (Whisper base.en through faster-whisper), the voice (Pocket TTS, 15 English voices) and turn taking run on the user's machine. The user can interrupt an answer by speaking. The model can use the agent tools with the same permissions, limited to the workspace folder by default. English only. The check in the real app ran on Windows.
- Computer control (since v3.0.6, beta): a model looks at the screen and uses mouse and keyboard in apps the user approves. Off by default. LU asks before it touches an app for the first time and before every action unless the session is set to autonomous. An action that sends, buys, deletes or changes permissions always needs a click and cannot be confirmed by voice. Terminals, password managers, system sign-in and permission dialogs, security settings, the Run dialog and windows with administrator rights are always off limits, and LU never types into a password field. A screen task ends after at most 25 steps or 8 minutes. Foreground only.
- Dictation: speech-to-text (Whisper). Read aloud: text-to-speech (Piper).
- 25+ built-in AI personas plus Caveman Mode (lite/full/ultra terseness)
- Thinking mode with universal tag stripping
- File upload with vision support
- Granular permission system across 7 tool categories
- Auto-update over a signed NSIS channel (Windows)
- 100% offline and private by default; AGPL-3.0 open source

## Supported Chat Models
- Qwen 3.8: 27B dense with vision through a separate mmproj file, plus a 9B distill for small cards (text only). Official, huihui abliterated and OrcaRouter uncensored builds, and two Ollama tags (v2.6.6, uncensored rows reworked in v2.6.8)
- Qwen 3-Coder: 30B and Next 80B
- Gemma 4: 31B / 26B MoE / E4B / E2B, native vision and tools, Apache 2.0
- GPT-OSS: 120B and 20B (via Ollama)
- GLM 4.7 Flash: 30B; GLM 5.1: 754B MoE
- DeepSeek R1: distilled 8B / 14B / 32B / Llama 70B
- Llama 4 Scout (16x17B MoE), Llama 3.3
- Mistral Small 3, Phi 4
- Abliterated/uncensored variants: Hermes 3, Llama 3.1 Abliterated, Qwen 3 14B Abliterated

## Supported Image Models (via ComfyUI)
- FLUX 2 Klein, FLUX.1 (8-10 GB VRAM)
- Z-Image Turbo, explicitly uncensored (10-16 GB VRAM)
- Juggernaut XL, Dreamshaper, RealVis (SDXL finetunes)
- ERNIE-Image (Baidu, Turbo and Base)
- Stable Diffusion XL, SD 3.5

## Supported Video Models (via ComfyUI)
- Wan 2.1 (1.3B and 14B), Wan 2.2
- HunyuanVideo 1.5
- LTX 2.3 (fast text-to-video on modest hardware)
- FramePack F1 (image-to-video on 6 GB VRAM)
- AnimateDiff Lightning, Mochi, Cosmos, SVD
- Removed in v2.5.9 because they could never run: CogVideoX (both bundles) and Pyramid Flow. Their pipelines were built against ComfyUI node names that no version of the wrappers registers, so every generation came back an error. Nothing already on disk is deleted; they are simply no longer offered. Wan 2.1 and 2.2, LTX, SVD, FramePack, Hunyuan, Mochi and Cosmos are unaffected.

## What Changed in v3.0.7

- Characters everywhere: a trained character can be picked on Image, Edit and every video tab of Create and sets the Z Image model it was trained for. A video with a character runs in two steps, a picture of the character first, then the clip from that picture. The chat's image and video tools take a character by name or trigger word.
- More ways to a character: frames from a video as a training set, a character trained on LU Cloud saved for local use, Edit with up to three of the photos a character was trained on, a text field per photo so a character can change clothes, and Low VRAM training for cards under 12 GB on Windows and Linux.
- ComfyUI workflows import in the format ComfyUI saves (Save or Export), from files or a whole folder, with subgraphs unpacked. Settings, AI Backends manages ComfyUI node packs on Windows and Linux (install from a repository link, update, remove to the trash). Every setting of an imported workflow can be changed in Create under Workflow parameters (issue 8, issue 14).
- Edit with Qwen-Image 2.1 takes up to ten images on your own machine, up from four.
- A chat can be duplicated or branched from any message, and Regenerate and Save & resend keep earlier answers as versions (issue 174). Editing an earlier message no longer sends it or deletes the rest (issue 172).
- Chat: sampling values per persona and group member (issue 175), fixes for personas (issue 170) and group chats (issues 165, 169), character cards as personas with their greeting, a ChatGPT export imported as chats, attached text files that fit the context (issue 171), reasoning effort for local models (issue 176), a think budget, a Chat Reply Timeout (issue 177), and Read aloud that speaks only what the answer shows (issue 160).
- ComfyUI in Create: a workflow imported in the Workflows window and paired with a Wan Animate model now runs in Motion Control. Wan Animate, Wan VACE and Wan S2V files show under Installed, and a search in Get new covers both catalogue tabs. A picture made in the chat loads through the app, so a current ComfyUI no longer shows it as a broken frame.
- Create: Enhance Image runs locally on Windows and Linux (Real-ESRGAN x4plus, 2x or 4x), the gallery has folders and shows the settings a result was made with, and the number of images per run sits next to the Create button.
- Models and backends: a LoRA from a CivitAI or Hugging Face link, Check files and Repair for installed models (issue 113), several OpenAI-compatible backends at the same time (issue 173), a GGUF file from anywhere on the drive, and context sizes taken from the server (issues 166, 167, 168).
- Mac (beta): training a character, Talking Character, Motion Control, Music with ACE-Step 1.5, Enhance Image with SeedVR2 3B, Remove Background through the subject detection built into macOS, Edit with a mask, HiRes fix, LoRAs and sizes for video, Extend Video, and a Qwen prompt enhancer.
- Fixes: ComfyUI starts when no Python on the machine has PyTorch, GPU layers -1 on AMD and Intel cards (issue 156), the microphone on Linux (issue 162). ComfyUI, LM Studio and the LU Engine answer to LU only when LU starts them.

## What Changed in v3.0.6

- Voice (beta): talk to LU and hear it answer. Runs locally, 15 English voices, tools by voice, a switch for thinking and a switch that unloads the voice engine.
- Computer control (beta): a model can work in the apps the user approves. Off until switched on, with per-app approval and a fixed list of apps that are off limits.
- The status line shows the LU mark with a plain word for what a run is doing, and the mark floats above the composer when the user scrolls up during a run.
- Dialogs have a new look.
- LU Cloud lists ten more chat models (beta) that their makers release to answer without refusing, marked Uncensored. The mark for measured models is now called No Filter.

## What Changed in v2.5.9 (26 July 2026)

A correctness release. Most of it is things that looked like they worked and did not.

- Security: two command-injection paths in the Coding Agent are closed. Shell quoting now follows the shell that actually runs (the old helper escaped for POSIX while the agent runs PowerShell on Windows), and PR links are validated against the characters GitHub itself permits. Both were reachable without a confirmation dialog, and both are pinned by tests.
- Coding agent: a surgical file_edit tool, num_ctx no longer pinned to 8192, verify loops no longer serve stale shell/test/read results from the same turn, context compaction no longer truncates freshly read code to 80 characters, diffs show deletions, .lurules loads, and sub-agents no longer collide on tool call ids.
- Remote access works with non-Ollama backends. The bridge used to forward everything to Ollama, so LM Studio, Lemonade and llama.cpp desktops answered a phone with an empty model list and a 400 on chat. It now translates between the phone's Ollama-shaped requests and an OpenAI-compatible backend.
- Local Create lanes report progress. The lanes themselves shipped in v2.5.8; what v2.5.9 fixes is installing them. Downloading a Motion Control, Music, Talking Character or Extend lane now names the file, counts it up, appears in the Downloads tray and can be cancelled. Interrupted model downloads resume from where they stopped instead of restarting at zero.
- A failed image is reported as a failure. When ComfyUI returned an error, the chat image tool passed the error text back as a result and told the model the picture was displayed; that no longer happens.
- Three video models removed (CogVideoX both, Pyramid Flow) because they could never run. Nothing in the catalogue is marked COMING SOON any more, because nothing was.
- Cloud renders are retained for seven days. Create states this once, with a download reminder. Trained characters on the user shelf are exempt from that clock. On the paid service, prompt prefixes can be cached briefly at the inference partner to speed up requests and are not stored permanently.
- Auto-approve now works on cloud models, read-aloud reaches the selected Piper voice, the image and video tools see GGUF models, and the Edit tab is labelled "Edit / Image to Image".

## How Locally Uncensored Compares
- vs Open WebUI: both are open-source Ollama frontends. Locally Uncensored adds image generation, video generation, an agent mode, mobile remote access and A/B compare. Open WebUI is chat-only.
- vs LM Studio: Locally Uncensored is open source (AGPL-3.0) and adds image, video, a coding agent and mobile access. LM Studio is closed-source and chat-only.
- vs Jan: Locally Uncensored is a lightweight Tauri app with built-in image and video generation. Jan is a polished Electron chat client with cloud API support.
- vs GPT4All: Locally Uncensored is an all-in-one creative suite. GPT4All is a popular local chatbot with document RAG.
- vs SillyTavern: Locally Uncensored has a modern Tauri UI, built-in image/video generation and 12-backend auto-detect. SillyTavern is roleplay-focused.
- vs KoboldCpp: Locally Uncensored offers full ComfyUI image/video generation. KoboldCpp is a strong creative-writing backend.
- vs Msty: Locally Uncensored is fully open source with creative AI. Msty is a proprietary multi-provider chat hub.

## Hardware Requirements
- Minimum: 8 GB RAM for text chat with small models (Llama 3.2 1B, Qwen 2.5 3B)
- Image generation: an NVIDIA card from 6 GB of VRAM with the starter checkpoint, 8 to 10 GB for FLUX
- Video generation: from 6 GB for AnimateDiff and image to video with FramePack F1, from 8 GB for the Wan 2.1 starter, 12 GB and up for Wan 2.2 and HunyuanVideo
- Best in class: 24 GB+ VRAM for 35B MoE and 35B+ chat models

## Installation
1. Download the installer for Windows, Linux or macOS from GitHub Releases (https://github.com/PurpleDoubleD/locally-uncensored/releases/latest). One-click install, no Docker, no terminal, no config files.
2. On first launch, the setup wizard scans for 12 local AI backends and offers one-click install links if none are running.
3. Pick a model from the Model Manager and start chatting. Switch to the Coding Agent for coding, or open Create to generate images and video. Add cloud providers in Settings to mix local and cloud.

## Privacy
- Local by default: chats, models and files stay on your machine
- No analytics kit and no crash reports
- On its own the app uses the network for two things: an update check against GitHub Releases shortly after start and at most every six hours after that, and a count of each press of the Cloud switch (which way, platform, app version, no identifier)
- A new version downloads in the background unless you switch that off under Settings, General, Updates; installing always needs a click
- While you are signed in to LU Cloud, in local mode too: an account check at lu-labs.ai every five minutes, and a press of the Cloud switch is counted for the account as well
- Everything else that goes online you start yourself: model and engine downloads, providers with your own API keys, web search in agent runs, and Remote Access
- Remote Access over the internet: Cloudflare relays the traffic, encrypted to and from Cloudflare, not end to end
- Cloud optional, two routes: bring your own key (your keys, your data), and the paid LU Labs Cloud service, where inference runs on our infrastructure partners
- App data and logs live in %LOCALAPPDATA%\lu-labs on Windows and ~/.local/share/lu-labs on Linux; the LU Engine chat models in %APPDATA%\Locally Uncensored\models and ~/.local/share/Locally Uncensored/models. Details: https://locallyuncensored.com/guide/storage/

## Links
- Homepage: https://locallyuncensored.com
- Handbook (install, first start, chat, agent, Code tab, Create, LU Cloud, settings, troubleshooting, FAQ): https://locallyuncensored.com/guide/
  - Handbook chapter, What it is: https://locallyuncensored.com/guide/what-it-is/
  - Handbook chapter, Install and update: https://locallyuncensored.com/guide/install/
  - Handbook chapter, First start (setup wizard): https://locallyuncensored.com/guide/first-start/
  - Handbook chapter, Chat: https://locallyuncensored.com/guide/chat/
  - Handbook chapter, Agent mode: https://locallyuncensored.com/guide/agent/
  - Handbook chapter, The Code tab: https://locallyuncensored.com/guide/code/
  - Handbook chapter, Create (images, video, audio, Character Studio): https://locallyuncensored.com/guide/create/
  - Handbook chapter, LU Cloud: https://locallyuncensored.com/guide/cloud/
  - Handbook chapter, Settings and troubleshooting: https://locallyuncensored.com/guide/settings-and-troubleshooting/
  - Handbook chapter, FAQ and glossary: https://locallyuncensored.com/guide/faq-and-glossary/
- Blog: https://locallyuncensored.com/blog/
- How to Run AI Locally (beginner guide): https://locallyuncensored.com/blog/how-to-run-ai-locally.html
- LM Studio Alternatives 2026: https://locallyuncensored.com/blog/lm-studio-alternatives.html
- Easiest Local AI Image Generator: https://locallyuncensored.com/blog/easiest-local-ai-image-generator.html
- Local AI on Your Phone: https://locallyuncensored.com/blog/local-ai-on-your-phone.html
- How to Run DeepSeek V4 Flash Locally: https://locallyuncensored.com/blog/how-to-run-deepseek-v4-flash-locally.html
- DeepSeek V4 Flash in the Cloud: https://locallyuncensored.com/blog/deepseek-v4-flash-cloud.html
- DeepSeek V4 Flash vs V4 Pro: https://locallyuncensored.com/blog/deepseek-v4-flash-vs-v4-pro.html
- Can You Run GLM-5.3 Locally: https://locallyuncensored.com/blog/run-glm-5-3-locally.html
- GLM-5.3 lokal nutzen (Deutsch): https://locallyuncensored.com/blog/glm-5-3-lokal-nutzen.html
- Можно ли запустить GLM-5.3 локально (Русский): https://locallyuncensored.com/blog/glm-5-3-lokalno.html
- Source code: https://github.com/PurpleDoubleD/locally-uncensored
- Discord: https://locallyuncensored.com/discord
