locally-uncensored

Locally Uncensored

The all-in-one local AI studio for your desktop. Chat, images, video and a coding agent in one app. Install it, pick a model, go.

Free and open source. In local mode everything runs on your own machine. No Docker, no terminal, no config files. The platforms are listed under Download.

Download · Three steps · What is in the box · How it compares · Models · FAQ

Chat that draws your images	Images and video in the same window
The coding agent shows the diff first	Agent mode does the legwork

---

Download

Take the latest build from Releases.

Platform	File	Status
Windows 10 and 11	Locally.Uncensored_3.0.10_x64-setup.exe (NSIS, recommended) or Locally.Uncensored_3.0.10_x64_en-US.msi	Tested every release, signed auto update channel
Linux	Locally.Uncensored_3.0.10_amd64.AppImage, Locally.Uncensored_3.0.10_amd64.deb or Locally.Uncensored-3.0.10-1.x86_64.rpm	Built on every release
macOS 14 or newer, Apple Silicon (Beta)	LU_3.0.9_aarch64.dmg	Mac: 3.0.10 follows shortly, this is the 3.0.9 build. Signed with an Apple Developer ID and notarized by Apple. Intel Macs are not supported. The app has run on one Mac so far. 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
Android 7.0 or newer (arm64)	LU-3.0.10-Android.apk	The phone app. Not in the Play Store yet: install the file from this page. SHA-256 dbd515d81bbc8ab70f21b850424f9f8b57ae88ae24962225beefc4116d139d2f

Some antivirus engines flag unsigned NSIS installers that download other binaries, which is a false positive. The installer is built by GitHub Actions from the public source on master, and the update channel is signed against a public minisign key, so you can verify both: see SECURITY.md.

Code signing for the Windows installers: free code signing provided by SignPath.io, certificate by SignPath Foundation. SignPath Foundation has accepted the project. The production certificate is not in place yet, and until it is the installers ship unsigned.

Current release: v3.0.10 (November 2026). Every change since 1.0.0 is in CHANGELOG.md.

3.0.10 fixes lost chats from 3.0.8 and 3.0.9. A backup no longer replaces the only complete copy of your chats, a note at launch says when the backup holds chats that are missing from your list, and Settings, Chat Backup has Restore chats from backup. The Android app is back as an APK, with the keyboard staying open in the Create prompt.

3.0.9 brings Liquid's LFM3 models for small machines: 1.7B and 2B, about 1 to 1.7 GB each, for text, pictures and speech, on the LU Engine, on Ollama or as MLX builds on a Mac; the LU Engine moves to llama.cpp b11960 for them. A group chat now sends every speaker the context you picked, can redo one speaker and lets you set who speaks when. A compaction is one quiet line in the chat, and the conversation before it stays one click away. Animate takes an end image and, with MiniMax H3, up to nine reference pictures, and Krea 2, Qwen-Image 2.1 and FLUX.2 take reference pictures for a style, a person or a pose. Local videos run up to 15 seconds without drifting in colour, and a video shows a rough preview while it renders. The trainer learns a style, a concept or a pose next to a character (Beta). "Use this model" on Hugging Face opens the download in LU, and a model from the search comes in a quant that fits your graphics card. A status menu sits in the Mac menu bar and the Windows and Linux tray, and the Linux AppImage starts again on systems without the tray library. What has only run in automated tests is listed under Known limits in the release notes.

3.0.8 makes web search in the chat work again: LU asks five search services in turn and needs no key for any of them. A video made in the chat plays in the chat, and a long one renders in the background while you write on. A picture attached in the chat reaches the image tool, and with two graphics cards chat and ComfyUI can each get their own. The LU Engine plans a model's memory from the file itself and offers a shorter context with one click. AMD cards on Windows get AMD's current ROCm build. New: the chat can read and change LU's settings with a confirmation card, the Model Manager can make a copy of a local model that refuses less, and Settings, Storage puts models, gallery and runtime in one folder of your choice. What has only run in automated tests is listed under Known limits in the CHANGELOG.

3.0.5 was about Create: LTX 2.5, FastH3 and YuE2 run on your own machine, Qwen-Image 2.1 gets a transparent background and its prompt enhancer, Edit, Remove Background and Enhance Image take up to 50 images in one go, and LU Cloud adds Qwen Image 2.1, LTX 2.5, MiniMax H3 and Wan 3.0 Reference. Chat takes any file, each model in a group chat can have its own persona, dropdowns stay inside the window (issue 149), and the speech install works again on Ubuntu 24.04.

3.0.4 fixed the Code agent and the agent in chat on LU Cloud: writing a long file no longer ends the run, "continue" picks up where a run stopped instead of starting over, the agent keeps a plan only when you ask for one, and cloud context is 32K by default and can be set per model. It also fixed the bugs reported on GitHub and Discord since 3.0.3 (issues 144 to 147): MiniMax H3 runs as a video model with sound, Edit takes up to three extra images with Qwen-Image 2.1, a missing encoder or VAE is named and offered as a download instead of ending in "Value not in list", the LoRA stack follows the loras folder, and group chat works with every LU Engine model.

3.0.3 was a hotfix: chats with many images no longer run the app out of memory, Create reads a model's family from the file instead of its name, and the five bugs reported on GitHub since 3.0.2 are fixed (issues 139 to 143).

3.0.2 was a hotfix: the Code tab lost its scrollbar and its input box once a conversation grew taller than the window (issue 138, a regression in 3.0.1) and is held to the window height again, and the count of open cloud models now takes the Create Studio endpoints in, which makes it fourteen video models and seven image models.

What is new in 3.0.1

• Qwen-Image 2.1 runs locally on Windows and Linux. One model generates a picture from a prompt and edits one from a reference image, with no mask to paint. About 16 GB of files, ComfyUI 0.37.0 or newer, and a research license that does not allow commercial use.
• Models has a LoRAs tab of its own. Search CivitAI, download into ComfyUI's models/loras folder, see what is installed with its size and trigger word, and delete what you no longer want.
• A GPU whose free memory could not be measured is planned more carefully, and the log now says total capacity instead of calling it free.
• The Linux AppImage stops handing its own runtime to every program LU starts, so a healthy system Python, pip, git or ffmpeg keeps working, and a distribution that blocks pip in the system Python keeps the isolated environment LU built.
• The local model runs one conversation at a time and says so, with a line showing how many chats are ahead. Stop works while a chat is still waiting.
• Temperature, Top P and Max tokens belong to the conversation you moved them in, instead of every open chat sharing one value.
• Cloud video renders offer the clip lengths the picked model really supports, read from the live catalog.

Mobile app (Android, iPhone)

LU is also available as an app for your phone. The current APK comes from the 3.0.10 release.

• Android: download LU-3.0.10-Android.apk from the 3.0.10 release. Needs Android 7.0 or newer on a 64-bit (arm64) phone. Allow installing apps from your browser or file app when Android asks. SHA-256: dbd515d81bbc8ab70f21b850424f9f8b57ae88ae24962225beefc4116d139d2f
• iPhone: in review at the App Store.
• Google Play: coming later.

What the app does:

• Run AI models on your phone and chat offline. Chats with on-device models stay on the phone.
• Ask questions about your own documents. Memory and personas.
• Optional LU Cloud with your account: larger hosted chat models, Create (images and video) and dictation.
• Connect to LU on your own computer over your home Wi-Fi and chat with the models on your PC.

Later APKs install over this one and keep your data. The Google Play version will be signed by Google: to switch to it, export your chats first (Settings, Data) and uninstall the APK.

Three steps

1. Install. Run the installer like any other program. On first launch a wizard looks for an AI engine that is already running on your machine, and installs one for you with a single click if there is none. It does the same for ComfyUI, which draws the images and video.
2. Pick a model. The model manager marks which models fit your hardware and downloads them in one click. It works with the engine you already have (Ollama, LM Studio and others) and can link models those tools already store, without copying them. Loading a model in Ollama or LM Studio can be cancelled, and an Ollama model that runs partly on the processor says so.
3. Create. Chat, Create, Code and Agent are tabs in one window. Type in one, switch to the next, keep the same models.

Walkthrough with screenshots: Getting started guide. New to local AI: the five minute beginner guide.

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
npm install
npm run dev          # browser dev mode
npm run tauri build  # desktop binary

setup.bat on Windows and setup.sh on Linux and macOS bootstrap Node, Git and Ollama for dev mode. See the contributing guide.

What is in the box

Chat. Models that answer directly, without refusals, next to the mainstream ones in the same menu. Thinking is shown as it happens, with an effort control on the models that offer one. Upload an image and ask about it (a local model with image input is downloaded together with its image part, the second file it needs to see pictures), chat with your own documents through a local index, talk instead of typing and have answers read back, keep a memory across conversations, switch personas, and import your ChatGPT, Claude or Gemini export so the old threads come with you. A long conversation folds its older turns into a summary instead of running out of room.

Create. Images and video on your own GPU through a ComfyUI the app installs, starts, repairs and updates for you. No node graphs. The lanes are text to image, edit and image to image with a mask for inpainting, remove background, video, animate an image, extend a clip, motion control, talking character, music, and Character Studio, which trains a character on your own card and drops it into your local LoRA folder. A LoRA picker with a stack and strength sliders reads the folder live, and everything you make lands in the gallery. Upscale and Erase Object are the two lanes that run in the cloud only. How it works.

Code. A coding agent that builds a map of the repo, edits only the lines you asked for, shows the diff before it applies anything, runs your tests and reads the failures, and carries typed git tools and per project rules in a .lurules file. Ask, Plan and Bypass modes per conversation, a file explorer with preview, and a workspace per project. A Local API in Settings puts every local model on one OpenAI compatible address behind a token, so your other coding tools can use the same machine.

Agent. Web search and fetch, reading and writing files, a shell, code execution, screenshots, image and video generation, and your own MCP servers on top. Long jobs go to background agents that work while you carry on, and a panel shows what is running. Every tool call passes a permission gate you control, and a read only run stays read only.

Voice (Beta). A spoken conversation with the chat model you picked. Speech recognition, the voice and the turn taking run on your own machine, in English, with fifteen voices. You can interrupt an answer by speaking, the model can use the agent tools with the same permissions, and you decide whether those tools reach your workspace folder or the whole computer. Voice runs in local mode only, and with the Cloud switch on it is hidden. Voice is in beta: it is new, and your computer may still show rough edges. Our check in the real app ran on Windows, and on a Mac Voice has been tried by hand in the app. VOICE.md has the setup and what was measured.

Computer control (Beta). A model looks at the screen and uses the mouse and the keyboard in apps you approve. It is off until you switch it on, asks once in every new chat before it starts, asks before every action unless you make a session autonomous, and never touches terminals, password managers or password fields. Anything that sends, buys or deletes needs your click, and so does a button without a name. In an autonomous session, moving the mouse pauses it, and a shortcut stops it at any time. It is in beta: it is new and not proven in every app yet, so watch what it does and stop it when it goes wrong. Our check in the real app ran on Windows. On a Mac it is in the app but not fully tested yet. COMPUTER-CONTROL.md lists what it asks and what it never touches.

Everywhere. Reach the app from your phone over your LAN or through a Cloudflare tunnel, paired by QR code and a passcode, off until you turn it on, with connected devices listed. Compare two models side by side, benchmark them on your own hardware for speed, cost and correctness, and take updates over a signed channel. Phone details.

Optional: LU Labs Cloud

Some models are larger than any desktop card. Flip the Cloud switch in the same app and those run on hosted GPUs instead, with the heavy Create lanes alongside them. The same account works in the browser at lu-labs.ai, on a plan or on credit packs that do not expire. The local app stays free either way, and switching back to local costs nothing: plans and prices. With Create in chat (Beta), a cloud chat model can put together an order for images, video, music or voice, and nothing runs until you approve it.

Since 3.0.0 the catalogue says what it can back up. Every cloud chat model was asked the same two questions twice and judged on the answer, and a longer three-round test on 2026-10-05 took the mark from five more: 19 of the 46 we measured answer in full, and only those carry the No Filter mark in the picker. Across all measured answers, 5% were refusals. Ten more chat models that their makers release to answer without refusing are listed as a group of their own, marked Uncensored (Beta). They were not part of the measurement above. On an active plan, 11 models on Hosted and 12 on Pro and Max cost no credits in chat, up to a daily limit. For images and video, 14 video and 7 image models run without a built-in content restriction, most of them with Uncensored in the name, and your account decides what they may produce.

How it compares

	Locally Uncensored	Open WebUI	LM Studio	Jan	SillyTavern
Chat	Yes	Yes	Yes	Yes	Yes
Image generation	Yes	No	No	No	Partial
Video generation	Yes	No	No	No	No
Image to image and image to video	Yes	No	No	No	No
Coding agent	Yes	No	No	No	No
Agent with tools and MCP	Yes	No	No	No	No
Works out of the box	Yes	No	Yes	Yes	No
Remote access from your phone	Yes	Partial	No	No	Partial
Compare and benchmark	Yes	No	No	No	No
Answers directly, without refusals	Yes	No	No	No	Partial
Voice in and out	Yes	Partial	No	No	Partial
Document chat	Yes	Yes	No	No	No
Runs without Docker	Yes	No	Yes	Yes	Yes
Open source	Yes	Yes	No	Yes	Yes

Deep dives: vs LM Studio · vs Jan · vs Open WebUI · vs GPT4All · vs Msty · vs KoboldCpp · vs SillyTavern · LM Studio alternatives · Best local AI apps 2026

Models that run well

The model manager holds the full catalog and marks what fits your machine. These are the ones worth starting with.

Text

Model	Download	Notes
Qwen 3.8 27B Uncensored	18 GB	Vision, tools and switchable thinking, 262K context. The IQ2_M build is 11 GB and fits a 12 GB card.
Gemma 4 12B Heretic	7.4 GB	Vision, a different voice from the Qwen line. The IQ4_XS build is for 8 GB cards.
Qwen3-VL 8B Abliterated	5 GB	Image understanding on an 8 GB card.
Hermes 3 Llama 3.1 8B	5 GB	Native tool calling. The small model to give the agent.
Llama 3.1 8B Abliterated	5 GB	Fast and reliable, the usual first download.
GLM 4.7 Flash Heretic	10 GB	30B class in an IQ2_M build that fits 12 GB.
Qwen 3.6 35B MoE Abliterated	24 GB	35B total with 3B active, vision and agentic coding.

GLM 5.3 is in the catalog too, but its smallest local quant is 217 GB, so on a desktop it is cloud territory. Kimi K3 is further out still.

Image

Model	VRAM	Notes
Juggernaut XL V9	6-8 GB	Photoreal SDXL, the friendliest entry point.
DreamShaper XL Turbo V2	6-8 GB	Anime and stylized.
FLUX.1 schnell or dev	8-10 GB	Fast, or slower and better.
FLUX 2 Klein 4B	8-10 GB	The newest FLUX, and the quickest of them.
Z-Image Turbo	10-16 GB	Unfiltered. Loads slowly below 16 GB.
Qwen-Image 2.1	16-24 GB	Generates and edits in one model. Needs ComfyUI 0.37.0 or newer. Research license, non-commercial.
Qwen-Image 2.1 (No Filter)	16-24 GB	The same model with a community text encoder that has its refusals removed (Heretic).
Noct Q (Qwen-Image 2.1, Unfiltered)	16-24 GB	A community finetune of Qwen-Image 2.1 with modified image model weights. Research license, non-commercial.
ERNIE-Image Turbo	24 GB	Baidu DiT, eight steps.

Video

Model	VRAM	Notes
AnimateDiff Lightning	6-8 GB	Four step animation, the cheapest way in.
FramePack F1	6-8 GB	Image to video on a small card.
Wan 2.1 1.3B	8-10 GB	Light text to video.
Wan 2.1 14B FP8	12+ GB	The quality step up.
Wan 2.2 TI2V 5B	12+ GB	Image and text to video in one model.
HunyuanVideo 1.5 FP8	12+ GB	Strong frame to frame consistency.
LTX Video 2.3 22B FP8	16+ GB	The newest LTX.

Music runs on ACE Step 1.5 Turbo at 6-8 GB, and Talking Character on Wan 2.2 S2V from 10-12 GB.

Bring your own engine

If you already run a local engine, the app finds it and lists it instead of installing a second one. It detects twelve of them: Ollama, LM Studio, Jan, GPT4All, KoboldCpp, llama.cpp, vLLM, LocalAI, text-generation-webui, TabbyAPI, Aphrodite and SGLang. It also ships its own LU Engine for people who have none. Cloud providers are optional and use your own keys: OpenAI, Anthropic, OpenRouter, Groq, Together, DeepSeek, Mistral, or any OpenAI compatible address you paste in.

FAQ

Is it really free and offline? Yes. AGPL-3.0, no account, no usage limits. In local mode chat, agent runs and image and video generation all happen on your machine, and after the model download they work without a connection. The app has no analytics kit and sends no crash reports. On its own it uses the network for two things: it checks GitHub Releases for updates shortly after start and at most every six hours after that, and it counts each press of the Cloud switch (which way, platform, app version, no identifier). While you are signed in to LU Cloud it also checks your account every five minutes. Settings, General, Privacy has the full list.

What does "uncensored" mean? Abliterated models have the trained in refusal behaviour removed from the weights, so it is not a jailbreak that can be patched or break. They answer directly, without refusals. Full guide.

What hardware do I need? Chat runs a 3B model on 8 GB of system RAM with no GPU at all, and an 8B model comfortably on a 6 GB card. Image generation starts at 6 to 8 GB of VRAM, and so does video on AnimateDiff Lightning or FramePack F1; the Wan and Hunyuan models want 12 GB or more. The model manager marks what fits before you download. Chat runs on NVIDIA, AMD and Intel cards through the Vulkan build of the LU Engine on Windows and Linux. Images and video run on NVIDIA, on AMD with ROCm on Linux (RX 6800 and newer, RX 7000, RX 9000) and, from 3.0.8, on Windows for RX 7000 and RX 9000.

Can I run it on a Linux server, without a screen? LU is a desktop app with a window, so it needs a graphical session: a desktop on the server, or one you reach over VNC or RDP. There is no --host flag. For other machines on your network, Settings, Voice & Remote, Local API with "Reachable on the network" gives every running local model an OpenAI compatible address on port 8129 (it binds all interfaces, needs a token, and serves the models that are loaded in LU, Ollama or LM Studio). "Start with LU" (3.0.8) brings it up on every launch. The full chat and agent window from another device is the Remote tab with LAN, on port 11435. On a server without a Vulkan driver the LU Engine runs on the processor; install mesa-vulkan-drivers (Debian, Ubuntu, Fedora) or vulkan-radeon (Arch) for an AMD card, and the app says so when it happens.

Can it replace ChatGPT or Claude? For most chat, writing and coding, yes, with a good 8B to 14B local model, and it stays private and unlimited. Frontier cloud models are still stronger on the hardest reasoning. You can add them with your own API keys, or use the Cloud switch.

Does remote access leak data? Remote access is off until you start it, sits behind a six-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.

What about macOS? The macOS row of the Download table says whether the current release has a Mac build. On a Mac, chat runs on local models through the bundled server with Metal, and Create makes images and video on the Mac itself, without ComfyUI, so what this README says about ComfyUI, VRAM and LoRA folders describes Windows and Linux. On any Mac you can use LU Cloud in the browser at lu-labs.ai, which runs chat, images and video on hosted GPUs.

Roadmap

Updated 5 November 2026. Everything up to 3.0.10 has shipped; Releases has the history.

Just out: 3.0.10

• Lost chats from 3.0.8 and 3.0.9: a note at launch when the backup still holds them, and Restore chats from backup in Settings
• The Android app as an APK again, with the keyboard staying open in the Create prompt

3.0.9

• Liquid LFM3 for small machines: text, pictures and speech, with llama.cpp b11960 in the LU Engine
• Group chat with the full context you picked, one speaker redone at a time, and a speaking order
• Animate with an end image and reference pictures, reference pictures in Create
• Local videos up to 15 seconds that keep their colours, and a preview while one renders
• Style, concept and pose training (Beta)
• "Use this model" from Hugging Face, and a quant that fits your graphics card
• A status menu in the menu bar and tray, and a Linux AppImage that starts again

Phone

• Android: the APK is available now, Google Play follows
• iPhone: the app is in review at the App Store

Soon

• 3D in Create: text to 3D and image to 3D
• LU Sync: your chats, memory and characters on all your devices
• Partner program
• Complete remote access via the mobile app: Chat, Agent, Code and all of Create from your phone, running on your PC

Later

• Ideas from the community, for example a roleplay mode. Add yours in Discussions or in #feature-requests on Discord

Last

• Our own model

Tech stack

Tauri v2 with a Rust backend, React 19, TypeScript, Tailwind CSS 4, Vite 8, ComfyUI for images and video, faster-whisper for speech to text, Piper for reading answers aloud, and Pipecat with Whisper and Pocket TTS for Voice.

Community

Discord: https://locallyuncensored.com/discord, with help channels for chat, images, video and the coding agent. Bugs and ideas: Issues and Discussions.

License

AGPL-3.0-only. See LICENSE.

---

In local mode your data stays on your machine.

Website · Beginner guide · Blog · Report a bug · Request a feature