koboldcpp.dev

Running KoboldCpp on NVIDIA GPUs, from Pascal to RTX 50

Most NVIDIA cards should use koboldcpp.exe (Windows) or koboldcpp-linux-x64, which include CUDA 12.1, with the CUDA backend and a current driver. Older cards or drivers use the -oldpc build with CUDA 11.4, which also needs no AVX2. If CUDA fails on a very old card, Vulkan is the fallback. The -nocuda file has no CUDA at all. We have no NVIDIA card, so the per-card rows below are upstream facts and user reports, not our tests.

Checked against KoboldCpp 1.124.2 · updated 13 Nov 2026
  • lab-lnx: Linux test VM · 4 vCPU · 15 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
How we test
On this page
  1. Which file for which NVIDIA card
  2. KoboldCpp CUDA 12.1 against CUDA 11.4 (oldpc)
  3. Old cards: Pascal, Maxwell and Kepler
  4. RTX 50-series
  5. Drivers
  6. CUDA or Vulkan on NVIDIA?
  7. NVIDIA-specific settings
  8. Check CUDA is in use

Download picker

Pick the file for your computer, with NVIDIA preselected.

Runs in your browser. Nothing is uploaded.

1. Your computer

Answer the questions to get your file.

Which file for which NVIDIA card

We have no NVIDIA card in the lab, so every row is a source, not a test. NVIDIA’s CUDA release notes say which cards a toolkit supports; the links go to GitHub issues from users.

NVIDIA generations and the KoboldCpp build to try
RTX 50koboldcpp.exe with CUDA; Vulkan if it hangsuntestedUsers report hangs and empty windows (issues 1611, 1616)
RTX 40, 30, 20koboldcpp.exe, CUDAuntestedThe main build is made for these (upstream)
GTX 16 (Turing), GTX 10 (Pascal)koboldcpp.exe, CUDA; then oldpc; then VulkanuntestedPascal problems reported on some versions (issue 1362)
GTX 900 (Maxwell)koboldcpp-oldpc.exe or VulkanuntestedArchitecture errors reported (issue 610)
Kepler (GTX 600/700, K80)koboldcpp-oldpc.exe, or VulkanuntestedK80 reports (issue 1409); CUDA 12 dropped Kepler
Datacenter: P40, V100koboldcpp-linux-x64, CUDAuntestedV100 report (issue 1390)

KoboldCpp CUDA 12.1 against CUDA 11.4 (oldpc)

What each NVIDIA-capable build contains (upstream)
koboldcpp.exe, koboldcpp-linux-x6412.1 + VulkanAVX2<Data path="assets.koboldcpp.exe.mb" unit="MB" /> (Windows)
koboldcpp-oldpc.exe, -linux-x64-oldpc11.4 + Vulkanno AVX2<Data path="assets.koboldcpp-oldpc.exe.mb" unit="MB" /> (Windows)
koboldcpp-nocuda.exe, -linux-x64-nocudanone; Vulkan and CPUAVX2<Data path="assets.koboldcpp-nocuda.exe.mb" unit="MB" /> (Windows)

The oldpc build always loads the no-AVX2 library, so on a modern CPU it generated 28.2 tokens per second against 47.4 for the nocuda file in our lab F03. Use it only for an old card or driver, or an old CPU.

Name history: before 1.93 koboldcpp.exe was CUDA 11 and koboldcpp_cu12.exe CUDA 12; since 1.94 plain koboldcpp.exe is CUDA 12 and CUDA 11 lives in -oldpc (upstream workflow files). A CUDA 13 Linux workflow exists, but no released file appeared on 13 November 2026 (unverified).

Old cards: Pascal, Maxwell and Kepler

We could not reproduce an architecture error without an old card. These are the messages users quote in issues, so the text is theirs, not ours:

Quoted from user reportsGitHub issues 610, 1362, 1409 · not reproduced by us
ggml-cuda was compiled without support for the current GPU architecture1CUDA kernel mul_mat_vec has no device code compatible with CUDA arch 5202CUDA error: no kernel image is available for execution on the device3
  1. ggml-cuda was compiled without support for the current GPU architectur: The build has no code for your card: try oldpc or Vulkan.
  2. CUDA kernel mul_mat_vec has no device code compatible with CUDA arch 5: Maxwell (arch 5.2).
  3. CUDA error: no kernel image is available for execution on the device: Same cause, other wording.

Fix, in order: update the driver, try koboldcpp-oldpc, then choose Vulkan in the Backend menu (backends).

RTX 50-series

Drivers

Windows: the driver from nvidia.com. Linux: the proprietary driver; nvidia-smi shows the driver version and the highest CUDA version it supports. CUDA 12 builds need a driver that supports CUDA 12 (NVIDIA’s compatibility table); an old driver gives an error at start. OSError: [WinError 1114] after a driver change means the GPU library failed to initialise: update the driver, and try -nocuda to find out whether CUDA is the cause (won’t start). We did not run these on an NVIDIA machine.

CUDA or Vulkan on NVIDIA?

CUDA is the normal choice. Vulkan is inside the same file, works on NVIDIA, and is the better pick when your card is too old for CUDA 12 or when you mix GPU vendors. We have no number for the speed difference on one card.

NVIDIA-specific settings

  • Use MMQ (on by default since 1.80; turns it off).
  • CUDA graphs are on for every architecture since 1.124.2 (release notes).
  • Windows shared memory: see the shared GPU memory trap.
  • Laptops with two GPUs: see detecting the GPU.

Check CUDA is in use

The log must show CUDA0 buffer lines. A CUDA file on a PC without an NVIDIA card does not fail: it runs on the CPU, 2 to 4 times slower to start F02, and the log can still say offloaded N/N layers to GPU F17. See how to check.

Frequently asked questions

Which KoboldCpp file for an NVIDIA GPU?

koboldcpp.exe on Windows or koboldcpp-linux-x64 on Linux, with the CUDA backend. Use the oldpc file only for an old card, an old driver, or a CPU without AVX2.

Does KoboldCpp support my GTX 1060 or P40?

Pascal cards are supported by CUDA 12, so the main build is the first thing to try; users have reported problems on some versions and the oldpc build or Vulkan are the fallbacks. We have no Pascal card, so we did not test it.

Does it work on RTX 50-series?

We could not test it. The CUDA 12.1 toolkit that the main builds use is older than the first toolkit for RTX 50 cards, and users have reported hangs and empty windows. Try Vulkan if CUDA fails.

CUDA or Vulkan?

CUDA is the normal choice on NVIDIA. Vulkan also works and is the fallback for cards CUDA does not support. We could not measure the two on one card.

What happened to koboldcpp_cu12.exe?

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 the CUDA 11 build is koboldcpp-oldpc.exe.

Is there a CUDA 13 build?

A CUDA 13 workflow exists upstream for Linux, but we did not see a released file on 13 November 2026, so treat it as unreleased.