KoboldCpp won't start or the window closes at once

Step 1: see the error before the window closes

A double-clicked program opens its own console and closes it when it stops, so the error vanishes with it. Starting it from a terminal keeps the text. The console is the log: KoboldCpp wrote no log file by default in our test F58, see where to find the log.

1. Open a terminal in the KoboldCpp folder

   WindowsLinuxmacOS

   Open the folder that holds koboldcpp.exe. Click the address bar, type cmd and press Enter. A Command Prompt opens in that folder. In PowerShell, type powershell in the address bar instead.

2. Run KoboldCpp from the terminal

   WindowsLinuxmacOS

   > koboldcpp.exe
   > koboldcpp.exe --model "D:\My Models\model.gguf"

   Put quotes around any path with spaces. If the file name is different (you renamed it, or it is the nocuda or oldpc file), type that name instead; see the download picker.

3. Check it worked

   You now see text that ends with either an error or a model-loading line. Match the last lines to a section below, or paste them into the analyzer above.

If you downloaded the file again because it did not work, make sure it came from GitHub: see fake sites.

“Failed to execute script” is not the real error

Users report this line on 63 issues. It is a wrapper from the launcher and only says that an exception was not handled. The cause is printed above it. In our busy-port test the real line was Cannot start the server: no socket could bind, and the wrapper came right after F49. Our missing-library tests ended the same way F56.

OSError: Cannot start the server: no socket could bind.1[PYI-10566:ERROR] Failed to execute script 'koboldcpp' due to unhandled exception!2

See Failed to execute script and why the last line is not the real error.

Old CPU without AVX2: “Illegal instruction” or an access violation

How to tell: the program stops right after Initializing dynamic library, with Illegal instruction or an access violation, on a PC that is more than about ten years old. Users report this on 10 issues (#56, #1350, #624). We have not reproduced the crash itself yet, because our test machines all have AVX2; the speed numbers below are measured. See Illegal instruction and access violation.

Fixes, in this order:

Relative differences only: this is a CPU lab with a small model F03 F15. The flags are described under --noavx2 and --failsafe.

DLL errors: PyInstallerImportError, WinError 1114, koboldcpp_cublas.dll not found

The single file unpacks its libraries into a temporary folder every time it starts. An error here means a library was removed, blocked, or could not load. Users report it on about 20 issues (#599, #1555, #705, #785, #778). Causes, most likely first:

1. Antivirus removed or blocked a file. See antivirus.
2. A stale or partial unpack. See the temp folder.
3. No usable NVIDIA driver for the CUDA build. In our CI runs the CUDA file fails with koboldcpp_cublas.dll (Windows) or libcuda.so.1 (Linux) not found when a CUDA flag is set and no NVIDIA driver is installed F56. Update the driver from nvidia.com.
4. To isolate CUDA, try koboldcpp-nocuda.exe. It runs on any PC F02 and can use Vulkan.

The matching rows: PyInstallerImportError, WinError 1114, cublas DLL not found. A missing koboldcpp_common.dll is the same kind of error: with --usevulkan and no graphics driver we saw it, so a missing or broken driver looks like this F57.

koboldcpp_hipblas.dll

This file belongs to the third-party ROCm build for AMD on Windows, a fork, not the official download. Use the official koboldcpp-nocuda.exe with Vulkan, or report the problem to the fork. See AMD.

Antivirus deleted or blocked it

Open Windows Security, then Protection history, and look for KoboldCpp or a file in its temp folder. Users report several detection names on #795, #779 and #2267. Before you restore anything, verify the file:

1. Make sure it came from the official GitHub release, not a lookalike site: is KoboldCpp safe?.
2. Compare its SHA-256 hash with the one in the release list F01.
3. Only then restore it, or exclude one folder that holds KoboldCpp.

Smart App Control blocks it (WinError 4551)

Smart App Control is a Windows 11 feature that blocks programs it does not trust. Users report WinError 4551 or a blocked-app message on #1946 and #2590 (November 2026). We could not reproduce it: Smart App Control only exists on clean consumer installs, which our test machines are not. Statement checked 14 Nov 2026; Windows changes it often.

Your options:

• Check that the file is the official one first, then see Windows Security and the Event Viewer entry for the blocked file.
• Turn Smart App Control off. Microsoft’s documentation says it cannot be turned back on without resetting Windows, so read the page before you do it.
• Run KoboldCpp on Linux instead, on the same PC.

See WinError 4551.

Stale temp folder (_MEI) or a full disk

The single-file build unpacks itself into a folder whose name starts with _MEI: under %TEMP% on Windows and /tmp on Linux. Our Linux log shows it: /tmp/_MEI77CIbl/kcpp_adapters/AutoGuess.json F18. A failed or killed run can leave one behind, and the error Failed to extract … decompression resulted in return code -1 points to a full disk or a damaged unpack (users report #2342). We have not measured the sizes yet on Windows.

1. Close KoboldCpp.
2. Free disk space on the drive that holds the temp folder. The CUDA file is 594 MB compressed and unpacks to more F01.
3. Delete only folders whose names start with _MEI in the temp folder, then start again.

See Failed to extract.

Windows 7 and other old systems

PrefetchVirtualMemory not found in KERNEL32.dll (#68) means Windows lacks a function the program uses. Upstream restored Windows 7 support in 1.43 (2023), but its FAQ lists it as not working and we cannot confirm the current state: we have no Windows 7 machine. Checked 14 Nov 2026.

Options: run Linux on the same PC, rent a cloud GPU or use Colab, or use Kobold Lite with the AI Horde (KoboldAI). Old builds may run, but they receive no security fixes, so be careful. See PrefetchVirtualMemory.

The launcher never opens

On Linux without a display, or over SSH, the launcher cannot draw its window: the messages are no $DISPLAY and Exception in Tkinter callback. Skip the launcher by giving a model on the command line: --model (and --skiplauncher where needed), or a saved .kcpps with --config. The config builder will write the command for you. Users report the Tkinter error on 27 issues. See no display and Tkinter callback.

NVIDIA driver or new GPU problems

The CUDA build needs a current NVIDIA driver; we have not verified the minimum version. Users report empty windows or hangs on RTX 50-series cards (#1611, #1616), which we have not tested. Old cards: using the GPU and NVIDIA. To isolate the cause, start koboldcpp-nocuda.exe and choose Vulkan.

A KoboldCpp crash after it starts

• While loading the model: the model won’t load or out of memory.
• When the context fills: an illegal memory access error, often with a quantised KV cache and ContextShift: out of memory.
• GGML_ASSERT or Segmentation fault: see the rows in the error list; --usemlock without mmap has been reported (#1323).
• After an update: go back to the previous release from the releases and compare; see how to update.

Not sure? Start the decision tree.