Locally Uncensored
An open-source local AI desktop application that supports running chat, image generation, video, voice, and computer-control models on the user's own machine, with cross-platform support for Windows, Linux, and macOS.
Timeline (last 90 days)
In the desktop app, users can use it by turning on the Cloud switch and selecting DeepSeek V4.1 Flash 1017 in Chat; the app pulls the catalog from the server and no client update is needed.
Using a rolling alias (such as deepseek-flash) cannot reproduce historical behavior, whereas a dated model ID (such as deepseek-v4.1-flash-1017) pins to a specific version and is suitable for change control and regression comparison in production.
Locally Uncensored 3.0.6 adds computer control that lets local models operate user-approved apps with the mouse and keyboard, with screenshots and clicks staying on the local machine and not uploaded to the cloud.
LU marks content read from the screen as untrusted, but the wording used to judge whether a button is sensitive comes from a list, and expressions the list does not recognize cannot be identified; the eight corrections and ten security fixes added after Windows verification have only been run in automated tests.
LU has no background mode on any platform; it works in the foreground using the user's real pointer.
On Linux and macOS, LU can only rely on screenshots and cannot determine what a click will press, so every click is asked about as a sensitive action; on Mac it also requires the two system permissions of Accessibility and Screen Recording, and the model must have image input capability.
LU's computer control has only been verified point by point on Windows (a Windows 10 machine); Linux has not actually been run in the app and requires an X11 session; macOS is integrated but not sufficiently tested; the control list and autonomous mode exist only on Windows.
Using voice commands, a local 9B model types notes in Notepad based on a control list, taking 11 to 22 seconds from the end of the voice command to the voice announcement of the result; this data was measured before the mechanism that asks on every new conversation went live.
The dedicated screen model Holo 3.1 9B is marked Beta in the catalog, with 16GB VRAM recommended as a starting point; in the author's Windows tests it completed a simple typing task in only one of three runs, and cannot coexist with a 9B chat model on a 12GB graphics card.
In the author's comparison of three runs each on two tasks, the runs with images totaled 497 seconds and the runs relying only on the control list totaled 381 seconds, all six completed, and relying on the list alone was actually faster.
29 more timeline events
All Facts (20)
In the desktop app, users can use it by turning on the Cloud switch and selecting DeepSeek V4.1 Flash 1017 in Chat; the app pulls the catalog from the server and no client update is needed.
50%UnverifiedUsing a rolling alias (such as deepseek-flash) cannot reproduce historical behavior, whereas a dated model ID (such as deepseek-v4.1-flash-1017) pins to a specific version and is suitable for change control and regression comparison in production.
50%UnverifiedLocally Uncensored 3.0.6 adds computer control that lets local models operate user-approved apps with the mouse and keyboard, with screenshots and clicks staying on the local machine and not uploaded to the cloud.
50%UnverifiedLU marks content read from the screen as untrusted, but the wording used to judge whether a button is sensitive comes from a list, and expressions the list does not recognize cannot be identified; the eight corrections and ten security fixes added after Windows verification have only been run in automated tests.
50%UnverifiedLU has no background mode on any platform; it works in the foreground using the user's real pointer.
50%UnverifiedOn Linux and macOS, LU can only rely on screenshots and cannot determine what a click will press, so every click is asked about as a sensitive action; on Mac it also requires the two system permissions of Accessibility and Screen Recording, and the model must have image input capability.
50%UnverifiedLU's computer control has only been verified point by point on Windows (a Windows 10 machine); Linux has not actually been run in the app and requires an X11 session; macOS is integrated but not sufficiently tested; the control list and autonomous mode exist only on Windows.
50%UnverifiedUsing voice commands, a local 9B model types notes in Notepad based on a control list, taking 11 to 22 seconds from the end of the voice command to the voice announcement of the result; this data was measured before the mechanism that asks on every new conversation went live.
50%UnverifiedThe dedicated screen model Holo 3.1 9B is marked Beta in the catalog, with 16GB VRAM recommended as a starting point; in the author's Windows tests it completed a simple typing task in only one of three runs, and cannot coexist with a 9B chat model on a 12GB graphics card.
50%UnverifiedIn the author's comparison of three runs each on two tasks, the runs with images totaled 497 seconds and the runs relying only on the control list totaled 381 seconds, all six completed, and relying on the list alone was actually faster.
50%UnverifiedOn Windows, LU reads the list of buttons and fields of the target window (UI Automation); models that can see images get screenshots plus the list, while models that cannot see images work from the list alone.
50%UnverifiedA screen task ends after at most 25 steps or 8 minutes, and the time the user spends answering questions counts toward it.
50%UnverifiedLU's pause shortcut is Ctrl+Alt+Space and its stop shortcut is Ctrl+Alt+Esc; on Mac these are Control+Option+Space and Control+Option+Esc; moving the mouse during an autonomous session also triggers a pause.
50%UnverifiedActions such as sending, purchasing, deleting, and changing permissions, as well as clicking unnamed buttons, links or menu items, are all treated by LU as sensitive actions that must be asked about in any mode.
50%UnverifiedAt no level can LU enter the terminal, password managers, system login and permission dialogs, security settings, the Run dialog, or windows with administrator privileges, nor will it type into password fields or accept pasting into password boxes.
50%UnverifiedThe computer control permission for each new conversation is valid only for the current conversation; a new conversation or each conversation after an app restart asks again, and no setting can skip this confirmation card.
50%UnverifiedComputer control is off by default under Agent Permissions in settings; once enabled, it still asks at four points: each new conversation, the first operation of each app, every action (supervised mode asks by default), and sensitive actions (always asked in any mode).
50%UnverifiedIn the Locally Uncensored app, image models require a minimum of 8GB of memory, and video models require a minimum of 16GB of memory.
50%UnverifiedThe Mac version of Locally Uncensored is currently in beta and has only been tested on one M5 Pro (48GB memory).
50%UnverifiedThe app marketplace catalog contains a total of 23 image models (10 of which are in the Unfiltered group) and 31 video models (14 of which are in the Unfiltered group).
50%