Use the app

The Bubo app prepares the files, starts the server and shows the requests and the log. The app starts the same bubo serve server as the command line. You can use the app, the command line, or both.

Start the app

1. Do the steps in Prepare the machine, and install the Microsoft Visual C++ Redistributable.
2. Extract the complete package into a folder. The folder name can have spaces and non-ASCII letters.
3. Double-click start.cmd in the package folder. It opens the app (bin\Bubo.Control.Web.exe).

If you prefer the command line, see Install and the quickstart.

First run

The app has seven pages: Server, Models, Connect, Requests, Logs, Settings and About. On the first run, the Server entry reads Get started until the engine and a model are installed. Get started has two steps and then you start the server.

1. Engine files. The download approval says “To download: 967.2 MiB” and “Needs 3.2 GiB (includes unpacking)”, with the free space of the drive. It also shows the licenses. Approve it, and the app downloads the files.
2. Model. Select the model. The default is Qwen3.8 Flash Next from Huihui. Each option shows its download size and the GPU memory it needs at 200,000 tokens, for example “Download 108.6 GiB · Needs about 91 GiB of GPU memory at 200,000 tokens”. The approval says “Needs 110.6 GiB (108.6 GiB + 2 GiB spare)” and shows the free space of the drive and the licenses. Approve it, and the app downloads the model. One progress line shows the downloaded size, the speed and the time left.
3. Start the server. On the Server page, select Start server. The memory check tells you if the setup fits (see Memory).
4. Connect a client. On the Connect page, copy the address and the key, or copy a snippet for your client (OpenAI Python, PowerShell, Anthropic Python or Pi). The base URL is http://127.0.0.1:8080/v1. The snippets use the model name that the server shows on the Server page: the package id, for example huihui-qwen38-flash-next-ud-q4-k-xl-v1. The name bubo also works.
5. Watch. The Requests page shows each request while it runs and after it ends (see Watch a request). The Logs page shows the log of the server.

Nothing downloads before you approve it. You can run the steps again at any time: a file that is already correct is not downloaded again.

On a 96 GB computer, the app preselects the lower-memory build of Qwen3.8 Flash Next (“Download 83.7 GiB · Needs about 66 GiB of GPU memory at 200,000 tokens”). The two full-size options stay in the list but cannot be selected. They say: “Does not fit this computer: needs about 91 GiB of GPU memory. Use the lower-memory model.” See Supported models.

Memory

Setup settings are on the Server page, in Server setup. The top row has Model, Context per session and Sessions. Beside Model are the link Add a model you have and the toggle Model file, which shows the full path. Under Context per session, the hint “Max 262,144” shows the largest context for this model, and Use puts that number in the field. For a plain model file (a .gguf you added), the rows MTP draft file (optional, faster writing) and mmproj file (optional, image input) follow. Each shows the chosen file name or None, with Browse… and, once a file is set, Clear. The ⓘ button beside a row explains it. A downloaded package brings its own helper files and shows only the KV cache row. KV cache has the choices f16 and Q8 (half the memory). The section Speed and behaviour has Prefill mode (Normal or Fast) and In-engine abliteration (Qwen3.8 Flash Next) (Off or On, off by default). The sections Network and Inference follow, with More inference settings under Inference. Each setting is one line: the name, the field, a short hint and an ⓘ button for more. At a window width of 700 px or more, the rows run in two columns, and a setting that you did not change shows Default. The helper file, KV cache, Prefill mode and In-engine abliteration settings are used after the server restarts. Select Save server setup to keep a change.

The memory and cache settings that you change rarely are in Settings, tab Advanced, section Memory and caches: Disk cache (off by default), Disk cache limit (default 8 GiB), Disk cache staging memory (default Automatic), RAM cache size, Allow disk cache and Image input (both On by default), and Image-token limit.

The memory check runs before the server starts. A setup that does not fit never starts. For a setup that fits only just, Server setup shows a box headed Tight. The server starts only after you tick the box: “Override: allow this tight setup. It may run out of memory, stop the server or freeze your computer.” When you have unsaved changes, the box is disabled and says “Save this setup before allowing the override.” Check again is below it.

Watch a request

While the server runs, the header of the Server page shows one line, for example “Writing · 54 tok/s”, “Reading prompt · 62%” or “Idle”. With more than one active session it reads, for example, “Session 2 · Writing · 54 tok/s · 3 active sessions”.

The Requests page has a table headed Live requests. It has these columns: Session, Phase, Progress, Live tok/s and Observed elapsed. Next to it, a chart shows the rate for the last 60 seconds.

• While Bubo reads a prompt, Phase is “Reading”. Progress shows the tokens read and the total, with the cached part and a percent. Live tok/s is the reading speed.
• While Bubo writes, Phase is “Writing”. Progress shows “N tokens written”. Live tok/s is the writing speed.
• A value that is not known shows “-”, never 0.

Under the live table, the charts of finished requests (Decode, Prefill and time to first token) stay on the Requests page. The Server page keeps the line “Last request” with Decode tok/s and Prefill tok/s. The command bubo status --watch shows the same live facts in a terminal.

Other models

The Models page shows the other models of the catalogue. See Supported models for the models that were tested.

Stop and close

• Stop and Restart are on the Server page. Stop is graceful. If the server has not stopped after 45 seconds, the app stops it by force.
• When you close the window, the server and the app stop. If you turn on Keep this app in the tray when closing in Settings, the window hides and the server keeps running. Select Quit in the tray menu to stop both.

Settings and data

Settings has these options for the app: Font (Serif or Sans serif), Check for updates at start, Start this app at login, Keep this app in the tray when closing (off by default) and Start the server with this app.

Model download folder

Settings, in the App section, has a row Model download folder. It shows the folder and the free space of its drive, for example “S: has 538.2 GiB free”. The row says: “New downloads go here. Models you already downloaded stay where they are and keep working.”

1. In Settings, select Change… and choose a folder.
2. The row shows the new folder at once.
3. Select Use default to go back to the default folder (models in %LOCALAPPDATA%\Bubo).

The download approval shows the folder and the free space, in the form “Downloads to FOLDER · S: has 538.2 GiB free”, with Change folder… beside it. Change folder… opens a folder picker (“Choose a folder for new model downloads”), saves the folder as the download folder (the same setting as above) and updates the space line at once. If you cancel the picker, nothing changes.

Settings also has a list Model folders with Add folder and Remove folder for extra folders. After you add a folder, the app lists the model packages in it, and you can select one for the Server page.

The app saves its data in %LOCALAPPDATA%\Bubo:

Folder or file	Content
gui\settings.json	The settings of the app
gui\logs\	The logs of the app and the updater
gui\webview2\	The browser data of the app
runtime\rocm-10.0.0-gfx1151\	The engine files
models\	The models

Update check

The app checks for a new version once, when it starts. The check is on by default. It contacts the Bubo release feed on GitHub and sends only the installed version number in its User-Agent. It downloads nothing until you choose to update on the About page. To turn the check off, clear Check for updates at start in Settings. The check and the download use the proxy from the environment (HTTPS_PROXY) or the manual proxy of the Windows Internet Options. Automatic proxy scripts (PAC) are not supported.

Update

When an update is available, About, Updates shows what the update downloads and what it keeps. For the update from 0.1.0 to 0.1.1 it says “Installed: 0.1.0 · Available: 0.1.1”, “Engine files: not redownloaded.”, “Models: not redownloaded.” and “Download size: 63.1 MiB.” The notes of the release are under a fold, headed “Update to 0.1.1”.

1. Select Download update (or Later). The app says “The update is downloaded and checked.”
2. Select Install and restart. Installing stops the server and restarts the app. Your models and settings stay.