FAQ and glossary

The questions that come up in the help channels, answered from the code, and a glossary of the words the app uses.

Questions from the help channels

Is it really free, and does anything leave my machine?

The desktop app is free under the AGPL-3.0 licence, with no account and no usage limit in local mode. Settings, General, Privacy states the app's own pledge: "100% local by default. Chat, agent runs, image & video generation all execute on your machine. No telemetry, no analytics, no model pings home. The only network calls LU makes unless you explicitly opt in are: update checks against GitHub Releases, cloud provider APIs (OpenAI, Anthropic, etc.) that you configure yourself with your own API keys, and one anonymous daily count to lu-labs.ai when you press the Cloud switch (which way it was pressed, platform, app version, nothing else)." Model downloads are network calls you start yourself. LU Cloud costs money and is optional.

What hardware do I need?

For local chat: a 7B model in the usual 4 bit quantisation is a 4 to 5 GB file and wants about 6 GB of card memory for a full GPU run; without a card it runs on the processor with 8 GB of RAM or more, slowly. For local images: a card with 6 to 8 GB of memory for the starter checkpoint. For local video: 6 to 8 GB for the lightest models, 12 GB or more for the Wan and Hunyuan models. For the local character trainer: 12 GB. The Models tab shows a coloured dot on every model that compares it with your card (chapter 3), and the setup card of every Create lane names its model and its memory need before you download. Everything that does not fit runs hosted on LU Cloud.

Does it run on a Mac?

No. There is no Mac build and there has never been one. The release workflow builds Windows and Linux. On a Mac, use the hosted studio in a browser at lu-labs.ai with a LU Cloud account.

Does it work on an AMD card?

Three answers, because three parts are involved.

Chat: the LU Engine is built with Vulkan on Windows and Linux, so it runs on any card with a Vulkan driver, AMD included. On Windows the app reads AMD card memory from the HIP SDK; on Linux it reads it without ROCm.
Images and video: ComfyUI needs a PyTorch that can drive the card. On Linux the app installs the ROCm build for the cards ROCm supports (RX 6800 and up, RX 7000, RX 9000). For RX 5000 cards and RX 6000 cards below the 6800 no official ROCm build carries kernels, and the app says so and installs the processor build instead, with a note about the community route (HSA_OVERRIDE_GFX_VERSION=10.3.0) that it does not set for you. On Windows the app installs AMD's own Windows ROCm build for RX 7000 and RX 9000 cards; the status line adds "Researched, not proven: LU owns no AMD card, so this path has never run on our hardware." Other AMD cards on Windows get the processor build.
Character training: NVIDIA, or Linux with a ROCm supported card. On Windows with an AMD card the trainer refuses before downloading anything and points at Cloud mode, which trains the same character.

The model answers with garbage on my small card.

Two known causes. A model under 7B, which the app warns about and which the wizard no longer offers. Or a card with very little memory (a 2 GB card was reported) where the layer split between card and processor went wrong; version 3.0.0 measures the layer count against the card, but the report is still open because that card was never seen in house. Set "GPU layers" to 0 under Settings, AI Backends, LU Engine (expert) to run on the processor; if the text is clean then, the card is the cause.

I have models in Ollama or LM Studio. Do I need to download them again?

No. The wizard and Settings, AI Backends, Providers connect to Ollama and LM Studio and their models appear in the picker. For GGUF files another tool stores, "Scan for local models" under Model Storage finds them and "Import" links them into the LU Engine folder without a copy, as long as both are on the same drive.

Where are my models, and how do I delete one?

LU Engine models are in %APPDATA%\Locally Uncensored\models on Windows and ~/.local/share/Locally Uncensored/models on Linux, or in the folder you set under Model Storage. Delete a model under Models, Installed with "Delete"; the box "Delete Model" shows the file path and the sentence "This removes the model file from your disk and frees the space." Ollama's models are Ollama's own and are deleted with ollama rm.

I downloaded a video model with the Get button and it is not in the picker.

This was reported in September 2026 for FramePack F1 and Wan 2.1 on a ComfyUI backend, and moving the files into a different ComfyUI folder by hand fixed it for that user. The report is still open. Until it is closed: ComfyUI must be running for the Image and Video rails to fill ("Start ComfyUI to see your video models"), and the files belong in the ComfyUI folder the app points at (Settings, AI Backends, ComfyUI, Path), under models/ in the subfolder the model type expects.

The installer is flagged by my antivirus. Is that safe?

Chapter 2 answers this at length. Short: the installer is not yet signed with a Microsoft certificate, the update channel is signed, and you can compare the SHA-256 with the one on the Releases page. Do not disable protection to install it; if the source and the checksum do not match, do not run it.

I closed the window and the GPU is still busy.

The X hides the app to the tray. After 30 seconds it unloads the local models, so the card is free. "Quit" in the tray menu ends the app at once.

Can I use it from my phone?

Not as an app on the phone. A paired phone can drive the app on your PC through the sidebar's "Remote" tab, over your network or a Cloudflare tunnel, with the four permission switches under Settings, Voice & Remote, Remote Access all off until you turn them on. The hosted studio at lu-labs.ai runs in a phone browser with a LU Cloud account.

Can other programs use my local models?

Yes. Settings, Voice & Remote, Local API starts an OpenAI compatible server on port 8129 for every model on this machine. It needs a token, which every request must carry, and "Reachable on the network" decides whether other computers on your LAN may use it.

Can I point it at my own server?

Yes. "Add Provider" under Settings, AI Backends, Providers with the preset "Custom (OpenAI-compat)" takes any address that speaks the OpenAI API. The app asks llama.cpp, vLLM, KoboldCpp and LM Studio servers what context they run with and shows it in the context dropdown with the label "from server".

Personas I made on my phone are not on the PC.

Persona sync is one way in 3.0.0, from the desktop to the paired device. The report is open. Write personas on the desktop under Settings, Agent, Personas.

How many tokens does a euro buy?

The handbook does not say, and the app does not either. Credits are spent at each model's own rate, the rates are on the pricing page, and the same wallet pays for images and clips at their own rates. Put a budget into the calculator on the pricing page to see what it buys in images or clips.

Is the agent sandbox a real sandbox?

It is a folder with a path jail, not a container. The dialog says so: "Workspace protection is a folder path jail, not a container or virtual machine. Commands run on this computer. Review tool requests before allowing them." The file tools cannot leave the folder; a shell command you approve can do anything your user can.

Is there a limit on how long I can chat?

Locally, none. In a conversation the limit is the context window; /compact folds older turns into a summary so the conversation continues. Hosted, the plan's credits and the Flash allowance are the limits, and the pricing page publishes the request counters too.

Glossary

Term	Meaning
Agent	A chat in which the model can call tools (files, shell, web, screenshots, image and video) and works towards a task step by step. Chapter 5.
Backend, provider, engine	The program that runs a model and answers the app's requests. The LU Engine is the built in one; Ollama, LM Studio and the others in the Providers list are alternatives; the cloud providers are remote services with your own key; LU Cloud is the hosted twin.
Checkpoint	A complete image model file, such as the Juggernaut XL starter checkpoint. ComfyUI loads one per render.
ComfyUI	The open source program that renders images, video and audio locally. The app installs, starts, repairs and updates it and builds its node graphs for you. Chapter 7.
Context window	How much text the model holds at once, in tokens. Chapter 4 explains the dropdown, the three labels and the memory cost.
Content policy	The Strict, Standard or Off setting for hosted images and video. It never applies to text or to anything local. Chapter 8.
Credits	The unit LU Cloud meters in. One shared wallet pays for chat, images and clips at each model's rate. Plans refill it monthly; packs add to it once and never expire.
CUDA, ROCm, Vulkan	Three ways software drives a graphics card. CUDA is NVIDIA's; ROCm is AMD's; Vulkan is a standard every vendor supports. The LU Engine uses Vulkan. ComfyUI's PyTorch uses CUDA on NVIDIA and ROCm on AMD where a build exists.
Embedding model	A small model that turns text into numbers so matching passages can be found. Document chat needs one (nomic-embed-text). It never writes an answer.
Flash	A class of hosted chat models that answer without credits on an active paid plan, inside the apps, up to a published daily token ceiling, one request at a time. Chapter 8.
GGUF	The file format of the models the LU Engine, Ollama and LM Studio run. A GGUF file carries the model and its metadata in one file. A GGUF that does not start with the GGUF marker is a broken download.
GPU layers, offload	A model is a stack of layers. The engine puts as many as fit on the card ("offload") and the rest on the processor. "GPU layers" under LU Engine (expert) overrides the choice: -1 all, 0 none, N a number.
KV cache	The memory where the model keeps the conversation while it works. It grows with the context length, which is why a large context needs a large card.
LoRA	A small add-on file that teaches an image model one thing, such as a character or a style, without replacing the model. Goes into ComfyUI's `models/loras` folder. The Character Studio trains one.
LU Cloud	The hosted twin of the app: the same window with the work on rented GPUs, paid in credits, also usable in a browser at lu-labs.ai. Chapter 8.
LU Engine	The built in llama.cpp server the app ships, listening on 127.0.0.1:8127 (embeddings on 8128). It runs GGUF models.
MCP	A standard for plugging an outside tool server into the agent. Settings, Agent, MCP Servers; the app starts them only through npx or uvx.
Memory	A small store of facts about you that the app injects into every prompt. The Memory toggle above the message box shows it.
Persona	A named system prompt. The default is "No Filter"; 25 are built in and you can add your own.
Quantisation (Q4_K_M, IQ2_M, fp8)	Shrinking a model's numbers to fewer bits so the file is smaller and fits a smaller card, at a small cost in quality. Q4_K_M is the usual choice for chat models; fp8 appears on image and video models.
RAG, document chat	Splitting your documents into pieces, finding the pieces that match a question with the embedding model, and handing them to the chat model. Chapter 4.
Sampling: Temperature, Top P, Top K, Max tokens	The four dials that shape how the model picks each word. Chapter 4 has the table.
Sandbox, workspace, path jail	The folder an agent may work in. The sandbox is a fresh folder under `~/agent-workspace`; a workspace is a project folder you picked. The path jail refuses file access outside it. Chapter 5.
Seed	The number that starts a render's randomness. The same prompt, settings and seed give the same picture; -1 means a random seed each time.
Steps, CFG, sampler	Render controls: how many refinement passes, how strongly the prompt steers, and which algorithm. The Quality presets set them; the advanced settings expose them.
Sub-agent	An agent started by the agent for a part of a task, bounded by the caps under Settings, Agent, Sub-agents.
Thinking, effort	A hidden draft a reasoning model writes before it answers, and the Low to Max control for how much of it. Chapter 4.
Token	The unit models read and write in, roughly three quarters of an English word. Context windows, the Flash allowance and credit rates are counted in tokens.
Trigger word	The word that summons a trained character in a prompt. It is also the character's file name.
VRAM	The memory on the graphics card. The thing that decides which models run fully on the card.

Previous chapter: Settings and troubleshooting. Back to the handbook overview.