diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 105d019..bd5b46e 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -27,14 +27,14 @@ jobs:
             cuda_repo_arch: x86_64
             cuda_package_suffix: "12-8"
             cuda_home: /usr/local/cuda-12.8
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
+            cuda_architectures: 75-virtual;86-real;90-real;120-real;120-virtual
             build_jobs: "2"
           - arch: aarch64
             runner: ubuntu-24.04-arm
             cuda_repo_arch: sbsa
             cuda_package_suffix: "12-8"
             cuda_home: /usr/local/cuda-12.8
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
+            cuda_architectures: 75-virtual;86-real;90-real;120-real;120-virtual
             build_jobs: "2"
     env:
       QWENTTS_CPP_BACKEND: cuda
diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
index 190e80a..6137e5c 100644
--- a/.github/workflows/wheels.yml
+++ b/.github/workflows/wheels.yml
@@ -33,7 +33,7 @@ jobs:
             cuda_package_suffix: "12-8"
             cuda_home: /usr/local/cuda-12.8
             wheel_build_tag: 1cu128
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
+            cuda_architectures: 75-virtual;86-real;90-real;120-real;120-virtual
             build_jobs: "2"
           - arch: aarch64
             runner: ubuntu-24.04-arm
@@ -42,7 +42,7 @@ jobs:
             cuda_package_suffix: "12-8"
             cuda_home: /usr/local/cuda-12.8
             wheel_build_tag: 1cu128
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
+            cuda_architectures: 75-virtual;86-real;90-real;120-real;120-virtual
             build_jobs: "2"
           - arch: x86_64
             runner: ubuntu-24.04
@@ -51,7 +51,7 @@ jobs:
             cuda_package_suffix: "13-0"
             cuda_home: /usr/local/cuda-13.0
             wheel_build_tag: 1cu130
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual;121-real;121-virtual
+            cuda_architectures: 75-virtual;80-real;86-real;90-real;121-real
             build_jobs: "2"
           - arch: aarch64
             runner: ubuntu-24.04-arm
@@ -60,7 +60,7 @@ jobs:
             cuda_package_suffix: "13-0"
             cuda_home: /usr/local/cuda-13.0
             wheel_build_tag: 1cu130
-            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual;121-real;121-virtual
+            cuda_architectures: 75-virtual;80-real;86-real;90-real;121-real
             build_jobs: "2"
     env:
       QWENTTS_CPP_BACKEND: cuda
diff --git a/AGENTS.md b/AGENTS.md
index 81e5071..dafbd32 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -12,9 +12,22 @@ The workflow runs on pushed tags that match `v*`, builds the public Linux CUDA
 through the configured `pypi` environment.
 
 The public `qwentts-cpp-python` package currently publishes CUDA 12.8 wheels
-only. CPU and CUDA 13 wheels are useful validation artifacts, but publishing
-multiple backend flavors under the same package name, version, and platform tag
-would leave pip with no reliable way to choose the intended runtime.
+only. Publishing multiple backend flavors under the same package name, version,
+and platform tag would leave pip with no reliable way to choose the intended
+runtime.
+
+The PR wheel workflow may still build CPU and CUDA 13 wheels as downloadable
+validation artifacts. Those artifacts are not uploaded to PyPI by the publish
+workflow.
+
+CUDA release wheels keep native cubins for sm_86, sm_90, and sm_120, plus PTX
+fallbacks for sm_75 and the newest supported CUDA architecture. This keeps the
+public CUDA 12.8 wheels under PyPI's default per-file upload limit while
+preserving broad GPU compatibility. Ada sm_89 GPUs can run the sm_86 cubin
+through CUDA's same-major binary compatibility. DGX Spark / GB10 sm_121 uses
+PTX fallback in the public CUDA 12.8 wheels. A CUDA 13 build with native sm_121
+can avoid that fallback, but it is not the PyPI default because the deployment
+server image currently targets CUDA 12.8.
 
 To prepare a release: