1LocalAITool v0.9.0 — a window that follows the system into dark mode, and an engine that says why it will not start

The window follows the system's light or dark setting. A new theme, *Match this computer*, wears Paper by day and Night shift once macOS, Windows or the Linux desktop goes dark, and changes the moment the system does. It is the default on a new install; a theme already chosen stays as it was.

Light or dark in one click. A switch at the bottom of the sidebar steps through *Match this computer*, Light and Dark, without a trip to *This computer*.

The first download can wait. The last setup step has *Skip the download*: nothing is fetched, and the app opens on *This computer › Models*, where a model you already have can be added, or one installed when you are ready.

Choose where models live before any arrive. The same step asks where to keep them, with *Browse…* and a button for each connected drive — so a Mac with a soldered drive can keep every model outside it from the start.

Download progress stays in sight. While a model downloads or moves, the sidebar shows how far along it is on every screen, and the model you pressed *Install* on carries its own progress bar.

Install is green, Remove is red. On *This computer › Models* the two read apart at a glance, and a pressed button stays pressed — *Removing…* — until the work is done, instead of springing back after a moment.

Moving the models no longer means typing a path. Under *Keep them somewhere else*, *Browse…* picks the folder and connected drives are one-click choices. The section says where the models are kept now, and after a move, where they went. Read-only drives, like a mounted installer, are not offered.

Paths are written the way the computer writes them. *Already on this computer* lists the exact folders it searches — C:\Users\…\.lmstudio\models on Windows, /home/… on Linux — and the example paths match the system. A file named only model.safetensors is listed under the name of the Hugging Face folder it came from.

A message that gets no answer is not counted. On the Free plan a chat message counts when its answer arrives, not when it is sent. A turn the engine never answered, or one stopped before it said anything, costs nothing.

What you asked can always be copied. Your own messages have a *Copy* button, shown without hovering when the reply failed, and the error offers *Copy my message*.

Logs and crash reports, in one place. An engine error, the chat banner and *This computer › Overview* open a page that shows where the logs are, with *Open the logs folder*, *Copy a report* for a support email, and how to catch a crash on each system — coredumpctl on Linux, Event Viewer on Windows, Console on a Mac. The app now keeps a log file of its own, so a packaged build leaves something to read after a failure.

A graphics card that fails hands over to the processor. When the engine dies on the card, the next start runs on the processor with the card's driver left out entirely, so a driver that crashes no longer leaves the computer without an engine. *Use the processor only* makes that the rule.

The error says what went wrong. After three failed starts the message quotes the engine's own complaint — or how it ended, when it printed nothing — and *Try again* starts it fresh instead of repeating the refusal. Chat waits through the short pause between attempts rather than giving up at the first one.

The graphics card gets what fits, not everything. The engine was told to put the whole model on the card, sized against the computer's memory; on a card with less memory than the computer, that failed at every start. It now loads as much as the card holds.

GPU compositing in the window is off. WebKit's GPU rendering is the likely cause of a toolbar that flickers and keystrokes that go missing while the engine works the card, so it is switched off on Linux. ONELOCALAITOOL_GPU_UI=1 turns it back on.

The .deb brings what the engine needs. It now depends on libgomp1, without which the engine cannot start, and libvulkan1.

“Use the app's folder” no longer loses models. With the models already in the app's own folder, moving them there again copied each file onto itself and then deleted it. The same folder now means there is nothing to move.

Graphics cards are found on current Windows 11. Detection relied on a tool recent Windows no longer installs, so the card went unseen. AMD cards now get the Vulkan build of the worker; they used to get the processor build while being told to use the card.

Laptops know when they are on battery. *Only when plugged in* now holds back on battery; a Windows laptop always read as plugged in.

Free space is measured where the models go. A full C: next to an empty D: no longer reads as room for a model that will not fit.

No more black windows flashing up whenever the app runs a helper program, and paths no longer show with a \\?\ prefix.

1lat works in a Windows terminal. Answers with accented or non-Latin text no longer stop halfway, it runs through the py launcher when python is only the Microsoft Store placeholder, its PATH hint is a PowerShell line, and 1lat batch-get --split writes ids like req:1 to files Windows accepts.