koboldcpp.dev

Download KoboldCpp: the right file for your computer

Download KoboldCpp only from its GitHub releases page. On Windows, pick koboldcpp.exe if you have an NVIDIA GPU, koboldcpp-nocuda.exe (86 MB) for AMD, Intel or no GPU, and koboldcpp-oldpc.exe only for CPUs without AVX2 or very old NVIDIA cards. Linux has the same three builds; Macs with Apple Silicon use koboldcpp-mac-arm64. Answer three questions below to get your exact file.

Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
  • lab-lnx: Linux test VM · 4 vCPU · 15 GB · no GPU
  • gha-w25: Windows Server 2025 test machine · 4 vCPU · 16 GB · no GPU
  • gha-w22: Windows Server 2022 test machine · 4 vCPU · 16 GB · no GPU
  • gha-u24: Ubuntu 24.04 test machine · 4 vCPU · 16 GB · no GPU
  • gha-mac15: macOS 15 test VM (Apple Silicon) · 3 vCPU · 7 GB · paravirtual Metal
How we test
1. Your computer

Answer the questions to get your file.

KoboldCpp download: all official files (version 1.124.2)

All official KoboldCpp files
SHA-256Direct link
koboldcpp.exeWindowsNVIDIA GPUs (also runs Vulkan and CPU)594CUDA 12.1 + Vulkan + CPUe306b58f9770… GitHub
koboldcpp-nocuda.exeWindowsAMD, Intel or no GPU85.7Vulkan + CPUc66550f4852a… GitHub
koboldcpp-oldpc.exeWindowsCPUs without AVX2, or old NVIDIA cards458.2CUDA 11.4 + Vulkan + no-AVX2 CPUf1820f7232c3… GitHub
koboldcpp-linux-x64Linux x64NVIDIA GPUs (also runs Vulkan and CPU)599.8CUDA 12.1 + Vulkan + CPUc66ae0106673… GitHub
koboldcpp-linux-x64-nocudaLinux x64AMD, Intel or no GPU102.4Vulkan + CPUe7bc9457030f… GitHub
koboldcpp-linux-x64-oldpcLinux x64CPUs without AVX2, or old NVIDIA cards454CUDA 11.4 build for older systems8fcfa63a355b… GitHub
koboldcpp-mac-arm64macOS (Apple Silicon)M-series Macs65.2Metal + CPUf6424d36081a… GitHub

What happens if you pick the “wrong” one

A wrong pick usually still runs. It costs download size, start time or speed:

What happens with another file
koboldcpp.exe (CUDA)A PC without an NVIDIA GPUWorks on the CPU. A 594 MB download, and ready in 11.7 s against 6.5 s for nocuda (Windows Server 2022). On Linux 9.1 s against 3.5 s.R-ci1-w22-cuda-base-q4_k_m
koboldcpp-oldpc.exeA CPU with AVX2Works, but always loads the no-AVX2 library: 28.2 against 47.4 tok/s, about 40% slower.R-ci1-w22-oldpc-base-q4_k_m
koboldcpp-nocuda.exeAn NVIDIA GPUWorks through Vulkan. CUDA is usually faster; we have no real NVIDIA GPU number yet.not measured
The CUDA file starts slowest on a PC without NVIDIAWindows Server 2022 runner (no GPU), 0.5B Q4_K_M model, KoboldCpp 1.124.2, 13 Nov 2026CPU lab, 0.5B model: relative differences only
  • koboldcpp-nocuda.exe6.5 s to ready
  • koboldcpp-oldpc.exe10.2 s to ready
  • koboldcpp.exe (CUDA)11.7 s to ready
Data
ItemValue (s to ready)
koboldcpp-nocuda.exe6.5
koboldcpp-oldpc.exe10.2
koboldcpp.exe (CUDA)11.7

Download CSV · Results: R-ci1-w22-cuda-base-q4_k_m, R-ci1-w22-oldpc-base-q4_k_m, R-ci1-w22-nocuda-base-q4_k_m

oldpc generates slowest on a modern CPUWindows Server 2022 runner (no GPU), 0.5B Q4_K_M model, KoboldCpp 1.124.2, 13 Nov 2026CPU lab, 0.5B model: relative differences only
  • koboldcpp-nocuda.exe47.4 tok/s
  • koboldcpp.exe (CUDA)42.3 tok/s
  • koboldcpp-oldpc.exe28.2 tok/s
Data
ItemValue (tok/s)
koboldcpp-nocuda.exe47.4
koboldcpp.exe (CUDA)42.3
koboldcpp-oldpc.exe28.2

Download CSV · Results: R-ci1-w22-cuda-base-q4_k_m, R-ci1-w22-oldpc-base-q4_k_m, R-ci1-w22-nocuda-base-q4_k_m

These are CPU lab numbers on one runner type, so read the differences, not the absolute speeds F02 F03.

Renamed files: cu12, nocuda, oldcpu

Old guides and videos use names from before version 1.94. Before 1.93, koboldcpp.exe was the CUDA 11 build and koboldcpp_cu12.exe the CUDA 12 one. Since 1.94, plain koboldcpp.exe is CUDA 12, and CUDA 11 lives in the oldpc file.

Old and current file names
koboldcpp_cu12.exekoboldcpp.exeCUDA 12. Before 1.93 plain koboldcpp.exe was CUDA 11.
koboldcpp_nocuda.exekoboldcpp-nocuda.exeUnderscores became hyphens.
koboldcpp_oldcpu.exekoboldcpp-oldpc.exeoldcpu became oldpc. It carries CUDA 11.
koboldcpp-linux-x64-cuda1210koboldcpp-linux-x64Linux CUDA 12 build.
koboldcpp-linux-x64-cuda1150koboldcpp-linux-x64-oldpcLinux CUDA 11 build.

The picker’s search box maps these names for you. The 1.93.2 release published old and new names side by side.

Files that don’t exist (and what to use instead)

  • No Intel Mac build. koboldcpp-mac-x64 returns 404 for this version. Build from source: build from source.
  • No Linux ARM build. koboldcpp-linux-arm64 returns 404. Build from source.
  • No Android APK. Use Termux: Android. A file that claims to be a KoboldCpp APK is not from the project.
  • No official Windows ROCm build. The AMD ROCm build for Windows is a third-party fork (YellowRoseCx) that lags behind official releases. See AMD.
  • No CUDA 13 build yet. A workflow exists upstream but we did not see a released file on 14 Nov 2026.

Check that your download is genuine

Compare the SHA-256 of your file with the one in the table above. A short version; the full guide is Is KoboldCpp safe?.

PowerShell
Get-FileHash .\koboldcpp-nocuda.exe -Algorithm SHA256

The result must match the table exactly. If it does not, delete the file and download it again from GitHub.

First launch: which backend to choose

The launcher’s Backend list shows only the libraries inside your file. The names are: Use CPU, Use CUDA, Use Vulkan, Use CPU (Old CPU), Use Vulkan (Old CPU), Use Vulkan (Older CPU) and Failsafe Mode (Older CPU), plus Use hipBLAS (ROCm) in the ROCm build (upstream, 1.124.2). Pick CUDA for NVIDIA, Vulkan for AMD and Intel, CPU if you have no GPU, and the Old CPU entries only for CPUs without AVX2. On a PC where nothing was chosen, the launcher showed “Use CPU” F30. See backends and not using the GPU.

Update KoboldCpp

KoboldCpp does not replace its own file. The launcher’s Update button opens the releases page to check for a newer version (upstream launcher text). To update:

  1. Download the new file from the release page, using the picker or the table above.
  2. Start it. Your saved .kcpps files may load in the new version; we have not tested loading across versions yet.
  3. Delete the old file once the new one works.

Releases come often: median 8 days between GitHub releases across the 137 we counted. Rolling builds (a pre-release called rolling on GitHub) carry the same seven file names and the newest fixes, for example for an unknown model architecture error: model won’t load.

Older versions

Open the GitHub releases list, choose the version, and expand its assets. Older versions use the naming of their time (see renamed files). Roll back when a new release breaks something for you, and compare with the release notes.

Frequently asked questions

Which KoboldCpp exe should I download?

With an NVIDIA GPU, koboldcpp.exe. With AMD, Intel or no GPU, koboldcpp-nocuda.exe. Only if your CPU has no AVX2 (older than about 2013) or you have a very old NVIDIA card, koboldcpp-oldpc.exe. The picker above chooses for you.

What is koboldcpp-nocuda?

The Windows build without the CUDA libraries. It has Vulkan and CPU support and is 85.7 MB instead of 594 MB. It runs on any PC, and Vulkan is the official way to use an AMD or Intel graphics card.

Is there a KoboldCpp APK for Android?

No. KoboldCpp publishes no APK, and any file that claims to be a KoboldCpp APK is not from the project. On Android you run KoboldCpp in Termux.

Do I need to install CUDA or Python?

No. The files bundle what they need. For the CUDA build you need a current NVIDIA graphics driver. We have not yet tested the exact minimum driver version.

How do I update KoboldCpp?

Download the new file from the GitHub release page and use it in place of the old one. The launcher's Update button only opens the releases page. KoboldCpp does not replace its own file.