Install

Supported hardware: Bubo is built and tested for AMD Strix Halo with 128 GB of memory on Windows, mainly laptops. It can work on other machines, with no promise.

96 GB machines: use a lower-memory model: the lower-memory build of Flash Next, one of the ISTA files IQ3_S, IQ3_XXS and IQ2_XS, or the Flash Next Coder. Set Variable Graphics Memory (AMD Software: Adrenalin Edition > Performance > Tuning) to the largest value offered (72 GB if available, at least 64 GB), then restart Windows. At 72 GB, Bubo uses the default 200,000-token context for one session with the lower-memory build. At 64 GB it does too when about 16 GB of system memory is free, and picks a smaller context when less is free. The ISTA files and the Coder need less GPU memory than the lower-memory build: about 64 GiB for IQ3_S, 57 for IQ3_XXS, 50 for IQ2_XS and 41 for the Coder at 200,000 tokens and one session, measured on a 128 GB machine. To save more, set KV cache to Q8 (Server page in the app, or --kv-cache-type q8_0): the context cache becomes 47 % smaller. These values are calculated: Bubo has not been tested on a 96 GB machine.

Use this setup for a Strix Halo machine. The speeds on this site come from a laptop with Radeon 8060S graphics at approximately 70 W. Bubo runs natively on Windows 11 x64 with AMD HIP.

Prepare the machine

The documented Qwen3.8 Flash Next setup uses 128 GB of system RAM. Deep context and images need more memory than short text requests.

1. Set GPU memory to 96 GB in the BIOS or the laptop control app.
2. Set the page file initial size to 131072 MB on the system drive.
3. Set the page file maximum size to 131072 MB.
4. Plug in the laptop’s own full-power adapter, then select the Performance profile of your maker (in Armoury Crate on ASUS laptops) and the Windows power mode Best performance. This gives about 70 W. See Power mode.
5. Install AMD graphics driver 32.0.31041.1004 (2026-08-16) or newer.
6. Install the Microsoft Visual C++ 2015-2022 Redistributable (x64). Most PCs already have it. Install it before you run bubo.exe: Windows stops bubo.exe at start if the runtime is missing. The start script checks for it. The Bubo package does not contain the Microsoft runtime files.

Windows 11 build 26200 is the tested operating system. The GPU memory setting can have the name “UMA frame buffer” or “carve-out”. A page file does not replace physical RAM for full-speed operation.

Get the package

1. Download bubo-v0.1.2-windows-x64-gfx1151.zip from the Bubo release page. The file is 71,949,037 bytes. Its SHA-256 value is f1a86f1d667c6fd2926f846c3b1cca9f7151791b599a07db9179275ad6117e89. After you extract it, bubo --version prints bubo 0.1.2 (88eec3ab1), engine gufo 0.26.0.
2. Extract the complete package into a folder.
3. Open PowerShell in the bin folder of the package. The bubo.exe file is in bin. The file start.cmd is in the package folder, one level up.

The release command is bubo. PowerShell uses .\bubo.exe when the executable is in the current folder.

Windows security prompts

Like many open-source Windows builds, Bubo v0.1.2 is not code-signed. Windows can show warnings.

• A zip that a browser downloaded carries a “downloaded from the internet” mark. Right-click the zip, select Properties, tick Unblock, select OK, then extract it.
• If SmartScreen shows “Windows protected your PC” for start.cmd, select More info, then Run anyway.
• If Smart App Control is on (common on new Windows 11 installs), Windows blocks unsigned apps and offers no per-app override. Bubo v0.1.2 does not run while Smart App Control is on.

Prepare the AMD runtime

The Bubo package carries no AMD files. The bubo prepare command downloads them.

.\bubo.exe prepare runtime

The command downloads these official AMD wheels from AMD’s ROCm index.

Component	Pinned version
rocm_sdk_core	10.0.0
rocm_sdk_libraries	10.0.0
rocm_sdk_device_gfx1151	10.0.0

The total download is 1,014,229,359 bytes (1.0 GB). Before the download, the command shows the address, exact size and full SHA-256 value of each wheel, the destination folders and the AMD license information. It then asks for your approval. If you decline, nothing is installed (exit code 3). The --yes option accepts the displayed plan. It prints the same text first. The command checks the SHA-256 value of each archive before it extracts any file. It installs 13 DLLs and the gfx1151 data files (about 318 MB) in the runtime folder. It saves the license texts in the licenses folder inside the runtime folder. The step needs no Python and no administrator access.

By default the runtime goes in the shared folder %LOCALAPPDATA%\Bubo\runtime\rocm-10.0.0-gfx1151. A new package folder therefore keeps the runtime. Use --runtime-dir PATH to choose another folder, and set BUBO_RUNTIME_DIR to the same path when you start the server. A runtime beside bubo.exe (the v0.1.0 layout) still works. The downloaded wheels stay in a cache folder (default ROOT\cache\rocm-10.0.0-gfx1151; change it with --cache-dir PATH).

If a download stops, run the same command again. The command resumes partial downloads that it started. Use --no-resume to start them again.

If your network needs a proxy, set HTTPS_PROXY before you run the command. Set NO_PROXY for addresses that must not use the proxy.

$env:HTTPS_PROXY = "http://proxy.example:8080"

AMD license terms

Bubo does not include AMD files. The runtime step of bubo prepare downloads AMD’s ROCm 10.0.0 packages for gfx1151 directly from stable.repo.amd.com. Before it downloads, it shows the address, size and SHA-256 value of each package, and the license information. AMD licenses ROCm per component, and components can have third-party terms. Read the AMD ROCm 10.0.0 license page and the license texts that Bubo saves in the licenses folder. Other AMD terms can apply to these files. If you continue, you accept the terms that apply.

The pinned runtime is ROCm 10.0.0 for gfx1151. ROCm 10.1.0 does not match the tested runtime files.

Get the model

Model files have separate licenses. Read the Qwen license before the download. The command also saves the license and notice files with the model.

.\bubo.exe prepare model

The command downloads eight pinned files and checks each one against its SHA-256 value. It builds nothing on your machine.

Files	Source	Size
Four UD-Q4_K_XL shards	Huihui GGUF repository (revision 73e9eb7c)	111,334,655,392 bytes (111 GB)
MTP predictor and vision projector	Unsloth base repository	3,694,111,200 bytes (3.7 GB)
Dense twin and Q4_K draft head	HaloForge Labs Hugging Face repository (commit bb72204d)	1,588,370,016 bytes (1.6 GB)

The total is 116,617,136,608 bytes (117 GB). Before it downloads, the command shows this plan and asks for approval. Use --yes to accept the plan.

The twin and the draft head are derived from the Huihui UD-Q4_K_XL shards. They match only that source model. Both use the Qwen Community License 1.0, not MIT. The source model is an abliterated model; read the usage warnings on its model card.

By default the files go in %LOCALAPPDATA%\Bubo. Use --output ROOT to choose another folder, for example .\bubo.exe prepare model --output "D:\BuboData". The model folder is ROOT\models\huihui-qwen38-flash-next-ud-q4-k-xl-v1\. The command writes bubo-package.json last. The server finds the other files through that file.

Plan about 121 GB of free space for the runtime and the model together (an estimate).

To prepare the runtime and the model with one approval, run .\bubo.exe prepare with no step name.

If a download stops, run the same command again. The command resumes. Use --no-resume to start again.

Use files you already have

1. Put all four shards in the model folder with the exact names from the command output.
2. Run .\bubo.exe prepare model --output ROOT. The command checks the files and downloads only what is missing.
3. Use --mtp-model PATH or --mmproj PATH to copy an existing MTP or vision file. The file must match the pinned hash.
4. Use --offline to use local files only. A missing file then fails before any network use (exit code 4).

If a file in the model folder does not match its pinned size or SHA-256 value, the plan lists it as replaces a file that did not match and adds its size to the download. Keep free space for the whole new file: the old copy stays until the new file passes the check. The command asks the usual question, “Download and prepare all files in this plan? [y/N]”, and --yes accepts it. Bubo downloads the new file and then replaces the damaged one. If you answer no, Bubo changes nothing (exit code 3).

Bubo does not replace a file that you supplied (for example with --mmproj), or a file whose real location is outside the model folder (a link or junction). The error names the file. Delete the file and run the command again. If a running Bubo server holds a file open, the command stops with exit code 7. Stop the server and run the command again.

The base Qwen UD-Q4_K_XL model also works with explicit file paths. Bubo then runs without the twin and the draft head. Use one source model at a time.

Model examination

The default checks find a wrong file, a wrong model variant and a truncated file. They use file sizes, the GGUF header and tensor table, and a quick fingerprint of the tensors. The server never reads every model byte at startup, including the first start.

To check every byte, add --verify to the prepare command. It forces a full SHA-256 check of the selected files. The server option --verify-model does the same for the eight prepared files before the server starts. On an SSD this takes about 2 to 4 minutes (an estimate).

Use the original (non-abliterated) model

The default model is an abliterated build. If you need the original Qwen3.8 Flash Next, prepare the original set. It uses the plain Unsloth shards with the base twin.

1. Make sure that you have about 119 GB of free disk space. The download is 116,617,136,000 bytes (108.608 GiB).
2. Run the prepare command.

.\bubo.exe prepare model --model-set original

Add --output ROOT to choose the parent folder, as for the default model. The command shows the plan first: the files, their sizes, the licence texts and the free space. Nothing downloads before you approve.

The set has eight files: the four UD-Q4_K_XL shards of unsloth/Qwen3.8-Flash-Next-GGUF (revision 766911a6), the MTP predictor, the vision projector, and the base twin and the draft head from HaloForgeLabs/bubo-qwen3.8-flash-next (revision 0a3c343b). The model licence is the Qwen Community License 1.0. The files go to ROOT\models\qwen38-flash-next-base-ud-q4-k-xl-v1\. The default ROOT is %LOCALAPPDATA%\Bubo.

3. Start the server with the manifest of the original set.

.\bubo.exe serve llm --manifest "$env:LOCALAPPDATA\Bubo\models\qwen38-flash-next-base-ud-q4-k-xl-v1\bubo-package.json" --served-model-name bubo

Preparing the original set does not change the saved default model. The default stays the model that you prepared first. You must pass --manifest each time you want the original set. bubo status shows the original set as not prepared, because status checks only the saved default.

The default profile applies on its own. To check that the server bound the twin and the draft head, add --print-effective-settings to the command. It prints twin_bound and draft_head_bound.

The base twin and the draft head work only with these shards. The Huihui twin does not fit them. The draft head is the same file for both models. On the original model, the twin and the draft head raised generation speed by 12.4 % (46.06 to 51.76 tok/s) in a test on an earlier build. See the model card for the quality checks.

Notices

1. Read NOTICE-windows.md in the package.
2. Read THIRD_PARTY_NOTICES.md.
3. Keep the licenses folder in the runtime folder.

Gufo and the pixmaate Windows layer retain the MIT license. AMD components and model files retain their separate licenses.

Continue with the quickstart.