Set up the AI

Open AI Settings in the sidebar. The AI here suggests color schemes (Suggest schemes on a house’s Schemes tab); the AI photo runs its own engine — see The AI photo. The page’s tabs are Providers & models, Routing by feature (see The AI features), Usage and AI engine console.

Above the tabs, a strip shows this computer: OS, CPU, Memory, GPU, Acceleration, and how the graphics card’s memory is used — VRAM used (Memory used where the graphics share the computer’s memory, as on a Mac), Free, LLM (the AI model loaded now) and, when other programs hold more than a little, Other apps with show apps: Using the GPU right now lists them. Copy debug info copies the computer’s AI state — hardware, driver, engine, memory, loaded models — to paste into a bug report.

Local · free and Online · metered switch the providers shown: the AI on this computer, or an online service with your own key.

Quick Setup

Run Quick Setup (on the Local · free side) sets up the AI on this computer: it looks at the computer, proposes a model that fits it, and applies the choice. It sets up the built-in llama.cpp provider only.

Check this PC’s real AI speed? — asked when no preset matches the computer. Run the 1-minute check downloads a small test model once (its size is shown) and measures how fast the computer really runs one; Skip — use estimates picks from the hardware’s specs instead.
The next screen shows what it Detected, the model it recommends (Default model), and what Apply setup will do: the model becomes the one every preset uses — except a preset you changed yourself under Routing by feature. When no local model can run well here, it says so: connect an online provider instead.
Apply setup downloads the llama.cpp engine (the program that runs models) and the model — several gigabytes, once each — with Cancel and Retry on each.
Then the speed settings. Some computers come pre-measured: Tuned settings for your hardware were applied ✓. Otherwise Optimize for this PC offers 2-minute optimize (tries the likely settings) or Full optimize (10 minutes or more, often much faster in the end); Skip keeps everything set up. While it measures it holds the graphics card and other AI features wait; closing Quick Setup then asks — Close and stop or Keep optimizing.

Once it’s set up, Re-run Quick Setup or Change model walk it again; nothing changes until you apply.

Where the files go. The engine and the models go in the app’s data folder (App Settings → Storage shows how much they take) — unless another app of the family on this computer already downloaded models and this one has none: then, from its first start, this app uses that app’s files instead of downloading them again. Quick Setup shows the choice at its top — Share …’s AI files, naming that app (picked), or Keep a separate copy for this app; a switch applies the next time the app starts. Nothing on disk is moved or deleted either way.

The offer. When you add your first house and no AI is set up yet, the app asks: Run Quick Setup, Connect an online provider or Skip for now. It asks once — again only after Reset Just House Paint. AI Settings is always there.

The models

The models are in the built-in provider’s Edit (the Built-in row under Local · free), in the Model Catalog. It offers three from the family’s tested list:

Gemma 4 26B-A4B (QAT) — the recommended all-rounder; runs on 8 GB graphics cards and up, with much of it in the computer’s own memory.
Gemma 4 12B (QAT) — the lighter, faster pick: runs fully on a 10–12 GB graphics card (tight on 8 GB) and needs little of the computer’s memory.
Gemma 4 E4B (QAT) — made for laptops with integrated graphics.

QAT: trained for the smaller size they run at. Each row says whether it fits this computer (Fits, Tight, CPU, Won’t fit), how fast it should run (fast, fine, slow, very slow), its Size on disk and what it needs (Needs … GB VRAM · … GB RAM); Recommended for this PC marks Quick Setup’s pick.

Set as default — every feature uses it, unless you changed its preset. It loads the first time it’s needed.
Load model / Unload model — load it now, so the first suggestion doesn’t wait; or free the graphics card’s memory (it loads again when next needed). Load now and Unload at the top do the same for the default model. The AI photo unloads it while it draws.
Under More actions: Tune & measure, Re-download, Open folder (in the desktop app), View on Hugging Face, Delete downloaded model (deletes the file; it stays listed and downloads again when needed) and Delete from catalog (what used it moves to another model).
Add model takes any GGUF model on Hugging Face: its repo, then Load model info from HF and a quant. Reset catalog brings back the built-in models and keeps yours.

Load the default local model into memory on startup (under the built-in provider, on at first) loads the model as the app starts: the start screen shows it loading — or Setting up the AI engine while the engine installs — with Continue without waiting. If it doesn’t load, What can I do? opens Troubleshooting.

Tune & measure (a model’s More actions) finds the fastest settings for this computer: Load & measure (then Measure again) measures its speed; Auto-tune tries real settings one after another (4 to 30 minutes) and fills the grid with the fastest — review it, then Apply (the model reloads). Remove applied config goes back to the defaults; Save for PC class offers the settings to every computer like this one.

The engine. The built-in provider’s Local engine card: Install engine when it’s missing, Uninstall engine (your models are kept), Reinstall (the same build again) and Update to … when a newer build is out (checked before it replaces yours). Details ▾ has View log (the engine’s own log), Acceleration backend where the card offers a choice, the Loaded models, and its settings: Models kept loaded at once (2), Unload an idle model after (seconds · 0 = never) (900 — a quarter of an hour), VRAM safety margin (MB), Default context cap (tokens), RAM headroom (MB), Speed bands (tokens/second) (what “fast”, “fine” and “slow” mean) and Engine updates (Notify or Off) — then Save — and Engine binaries (where each build downloads from).

For experts: PC class configs… (launch settings shared by computers with the same memory) and Global launch defaults… (the switches under every model’s settings), in the built-in provider’s Edit; Reasoning levels… (what each thinking level asks a provider for), in a saved provider’s Edit.

Your graphics hardware changed — said once when the computer’s graphics card is different: a different model may now fit; Run Quick Setup checks again.

Online providers

Add provider under Online · metered: its Provider type — OpenAI, Anthropic (Claude), Gemini (Google), OpenRouter, DeepSeek, xAI (Grok), Mistral, or any OpenAI-compatible service — its Base URL, API key and Default chat model (Fetch lists its models). Test connection checks it before you Save provider; the saved key is never shown again (leave the box blank to keep it). On the list, Test checks a provider again (Connected or Failed).

Set as default makes the features run on it. Presets you changed by hand stay as they are unless you tick Also overwrite presets I customized.

What goes online: the house’s colors as words and numbers, which parts are painted, and the words you add — never the photo. The key is kept in the app’s data and goes into a backup.

Usage

Usage counts every AI call — calls, tokens (prompt and completion) and the est. cost — By feature and By provider. A model on this computer costs $0. Cloud pricing sets the price of an online model (US dollars per million tokens), used for the next calls’ cost. Reset ledger starts the counts over.

The AI engine console

AI engine console shows two live logs: Server log — the app’s own, the same as App Settings → Logs — and Engine output, the llama.cpp engine’s own words (why a model didn’t load, say). Info + and Warnings + filter the lines; Pause stops the scrolling and Resume follows again.

The AI tasks panel

Long work shows here: finding a house’s parts, an AI photo, installing the part finder or the AI photo, loading a brand’s colors, the AI’s schemes, and tests in AI Settings. Open it with the sparkle button at the title bar’s right, or AI tasks in the sidebar. The button shows how many tasks run; a red count means one failed (“1 AI task failed — open the queue”).

Running — each task with its progress and Cancel; Cancel all when more than one runs. An AI answer shows how far it is (reading the prompt, the first words, the tokens) — stalling · … when it’s slower than the model’s usual pace, stuck · … when nothing new has come for a long while. Show preview shows the answer as it arrives.
Recent — what finished, how long it took and how it ended. A failed task stays with its reason until you Dismiss it (the ×). Clear empties the list; Show all shows the older ones.

While you’re on another screen, a running task also shows at the bottom right of the window, with its progress, Cancel and Details.