Skip to content

Python 3.11 v0.3.20-cu126-cp311-win

Latest

Choose a tag to compare

Python 3.11, Windows CUDA 12.6.3 wheel, built 2026-04-20 against llama.cpp @ 227ed28e128e.

CUDA Toolkit isn't required — the CUDA 12.6 runtime is bundled inside the wheel. A local CUDA Toolkit installation won't interfere; you only need a current NVIDIA driver.

Install

ℹ️ CUDA Toolkit isn't required. The wheel ships with the CUDA 12.6 runtime DLLs it needs (cudart, cublas, cublasLt, etc.) vendored in. A local CUDA Toolkit installation won't interfere — you only need an NVIDIA driver supporting CUDA 12.6 (any recent GeForce/Quadro driver; check nvidia-smi shows CUDA Version: 12.6 or higher).

  1. Verify your NVIDIA driver is >= 561.17 (nvidia-smi).
  2. Download llama_cpp_python-0.3.20+cu126.ll227ed28e128e-cp311-cp311-win_amd64.whl below.
  3. pip install path/to/llama_cpp_python-0.3.20+cu126.ll227ed28e128e-cp311-cp311-win_amd64.whl.
  4. Verify: python -c "from llama_cpp import Llama; print('ok')".

Build forensics

Property Value
llama.cpp SHA 227ed28e128e
Base version 0.3.20
CUDA version 12.6.3
MSVC toolset 14.44
Python 3.11
Wheel size 356871737 bytes (340.3 MiB)
Wheel SHA-256 5b52775de9ad1e14e11e8b3b6cd2a1ce18a6b39222d1e20f6a2568e1f2a7da34
Build date (UTC) 2026-04-20
Dispatch run https://github.com/ThongvanAlexis/llama-cpp-python/actions/runs/24656443278
Smoke-test green (publish gated on needs: [build, smoke-test])