diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 720f8ae..e72388f 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -461,17 +461,58 @@ jobs:
       # job *after* the bundle was uploaded (and typically published server-side), while
       # the signed jars + .asc files already exist in each module's target/ (signing happens at
       # verify). Collecting on failure lets the github-snapshot job still attach them.
-      - name: Collect signed artifacts
+      # NOTE: the fat jar (jar-with-dependencies) is EXCLUDED here — it is no longer attached/
+      # deployed to Central (srcmorph-cli/pom.xml sets <attach>false</attach>), so it is unsigned
+      # in target/ at this point. It is built + signed + collected by the next step instead. This
+      # step runs BEFORE that one so the deploy-signed plain/sources/javadoc jars are captured
+      # before the classifier loop rebuilds (and clobbers) srcmorph-cli/target.
+      - name: Collect signed Central artifacts (plain/sources/javadoc; fat jars handled next)
         if: ${{ !cancelled() }}
-        # All three reactor modules produce their own jars/asc files (root pom has no jar).
         run: |
           mkdir -p signed-snapshot-assets
-          cp srcmorph/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
-          cp srcmorph/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
-          cp srcmorph-cli/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
-          cp srcmorph-cli/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
-          cp srcmorph-maven-plugin/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
-          cp srcmorph-maven-plugin/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
+          for m in srcmorph srcmorph-cli srcmorph-maven-plugin; do
+            for f in "$m"/target/*.jar "$m"/target/*.jar.asc; do
+              [ -e "$f" ] || continue
+              case "$f" in *-jar-with-dependencies.jar|*-jar-with-dependencies.jar.asc) continue ;; esac
+              cp "$f" signed-snapshot-assets/
+            done
+          done
+      # Build one srcmorph-cli fat jar per net.ladenthin:llama native classifier + the default
+      # all-platform CPU fat jar, GPG-sign each, and add them to the GitHub-Release assets. These
+      # are GitHub-only (never Maven Central): each bundles exactly one llama native binary, so a
+      # GPU jar runs only on its own OS/arch/backend, while the default (no classifier) bundles the
+      # multi-platform CPU natives and runs anywhere. Signing uses the key setup-java already
+      # imported (armored detached .asc, same shape Central produces).
+      - name: Build & sign classifier fat jars (GitHub-only assets, never Central)
+        if: ${{ !cancelled() }}
+        env:
+          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+        run: |
+          set -euo pipefail
+          if [ -n "${MAVEN_GPG_PASSPHRASE:-}" ]; then echo "::add-mask::${MAVEN_GPG_PASSPHRASE}"; fi
+          # Source of truth: the <classifier> set in java-llama.cpp llama/pom.xml for the pinned
+          # ${llama.version} (all published to Maven Central). Keep in sync on a llama bump.
+          CLASSIFIERS=(
+            cuda13-linux-x86-64 cuda13-windows-x86-64
+            vulkan-linux-x86-64 vulkan-linux-aarch64 vulkan-windows-x86-64
+            opencl-windows-x86-64 opencl-windows-aarch64 opencl-android-aarch64
+            rocm-linux-x86-64 rocm-windows-x86-64
+            sycl-fp16-linux-x86-64 sycl-fp32-linux-x86-64 sycl-windows-x86-64
+            openvino-linux-x86-64 openvino-windows-x86-64
+            msvc-windows
+          )
+          build() { mvn -B --no-transfer-progress -pl srcmorph-cli -am -DskipTests -Dmaven.javadoc.skip=true -Dllama.classifier="$1" package; }
+          for c in "${CLASSIFIERS[@]}"; do
+            build "$c"
+            f=$(ls srcmorph-cli/target/srcmorph-cli-*-jar-with-dependencies.jar)
+            cp "$f" "${f%.jar}-${c}.jar"
+          done
+          # default CPU build LAST so the unsuffixed fat jar is the all-platform CPU variant
+          build ""
+          for f in srcmorph-cli/target/srcmorph-cli-*-jar-with-dependencies*.jar; do
+            printf '%s' "$MAVEN_GPG_PASSPHRASE" | gpg --batch --yes --pinentry-mode loopback --passphrase-fd 0 --detach-sign --armor "$f"
+            cp "$f" "$f".asc signed-snapshot-assets/
+          done
       - uses: actions/upload-artifact@v7
         if: ${{ !cancelled() }}
         with:
@@ -542,17 +583,58 @@ jobs:
       # job *after* the bundle was uploaded (and typically published server-side), while
       # the signed jars + .asc files already exist in each module's target/ (signing happens at
       # verify). Collecting on failure lets the github-release job still attach them.
-      - name: Collect signed artifacts
+      # NOTE: the fat jar (jar-with-dependencies) is EXCLUDED here — it is no longer attached/
+      # deployed to Central (srcmorph-cli/pom.xml sets <attach>false</attach>), so it is unsigned
+      # in target/ at this point. It is built + signed + collected by the next step instead. This
+      # step runs BEFORE that one so the deploy-signed plain/sources/javadoc jars are captured
+      # before the classifier loop rebuilds (and clobbers) srcmorph-cli/target.
+      - name: Collect signed Central artifacts (plain/sources/javadoc; fat jars handled next)
         if: ${{ !cancelled() }}
-        # All three reactor modules produce their own jars/asc files (root pom has no jar).
         run: |
           mkdir -p signed-release-assets
-          cp srcmorph/target/*.jar signed-release-assets/ 2>/dev/null || true
-          cp srcmorph/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
-          cp srcmorph-cli/target/*.jar signed-release-assets/ 2>/dev/null || true
-          cp srcmorph-cli/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
-          cp srcmorph-maven-plugin/target/*.jar signed-release-assets/ 2>/dev/null || true
-          cp srcmorph-maven-plugin/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
+          for m in srcmorph srcmorph-cli srcmorph-maven-plugin; do
+            for f in "$m"/target/*.jar "$m"/target/*.jar.asc; do
+              [ -e "$f" ] || continue
+              case "$f" in *-jar-with-dependencies.jar|*-jar-with-dependencies.jar.asc) continue ;; esac
+              cp "$f" signed-release-assets/
+            done
+          done
+      # Build one srcmorph-cli fat jar per net.ladenthin:llama native classifier + the default
+      # all-platform CPU fat jar, GPG-sign each, and add them to the GitHub-Release assets. These
+      # are GitHub-only (never Maven Central): each bundles exactly one llama native binary, so a
+      # GPU jar runs only on its own OS/arch/backend, while the default (no classifier) bundles the
+      # multi-platform CPU natives and runs anywhere. Signing uses the key setup-java already
+      # imported (armored detached .asc, same shape Central produces).
+      - name: Build & sign classifier fat jars (GitHub-only assets, never Central)
+        if: ${{ !cancelled() }}
+        env:
+          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+        run: |
+          set -euo pipefail
+          if [ -n "${MAVEN_GPG_PASSPHRASE:-}" ]; then echo "::add-mask::${MAVEN_GPG_PASSPHRASE}"; fi
+          # Source of truth: the <classifier> set in java-llama.cpp llama/pom.xml for the pinned
+          # ${llama.version} (all published to Maven Central). Keep in sync on a llama bump.
+          CLASSIFIERS=(
+            cuda13-linux-x86-64 cuda13-windows-x86-64
+            vulkan-linux-x86-64 vulkan-linux-aarch64 vulkan-windows-x86-64
+            opencl-windows-x86-64 opencl-windows-aarch64 opencl-android-aarch64
+            rocm-linux-x86-64 rocm-windows-x86-64
+            sycl-fp16-linux-x86-64 sycl-fp32-linux-x86-64 sycl-windows-x86-64
+            openvino-linux-x86-64 openvino-windows-x86-64
+            msvc-windows
+          )
+          build() { mvn -B --no-transfer-progress -pl srcmorph-cli -am -DskipTests -Dmaven.javadoc.skip=true -Dllama.classifier="$1" package; }
+          for c in "${CLASSIFIERS[@]}"; do
+            build "$c"
+            f=$(ls srcmorph-cli/target/srcmorph-cli-*-jar-with-dependencies.jar)
+            cp "$f" "${f%.jar}-${c}.jar"
+          done
+          # default CPU build LAST so the unsuffixed fat jar is the all-platform CPU variant
+          build ""
+          for f in srcmorph-cli/target/srcmorph-cli-*-jar-with-dependencies*.jar; do
+            printf '%s' "$MAVEN_GPG_PASSPHRASE" | gpg --batch --yes --pinentry-mode loopback --passphrase-fd 0 --detach-sign --armor "$f"
+            cp "$f" "$f".asc signed-release-assets/
+          done
       - uses: actions/upload-artifact@v7
         if: ${{ !cancelled() }}
         with:
diff --git a/srcmorph-cli/README.md b/srcmorph-cli/README.md
index 29874ca..0213369 100644
--- a/srcmorph-cli/README.md
+++ b/srcmorph-cli/README.md
@@ -14,7 +14,19 @@ java -jar srcmorph-cli-<version>-jar-with-dependencies.jar <path-to-config.json-
 ```
 
 The fat jar (`srcmorph-cli-<version>-jar-with-dependencies.jar`, built by `mvn package`) bundles every
-dependency, including a logback binding, so it runs standalone. Both `.json`/`.js` (parsed via a
+dependency, including a logback binding, so it runs standalone.
+
+**Download:** the pre-built fat jars are attached to each
+[GitHub Release](https://github.com/bernardladenthin/srcmorph/releases) (with a `.asc` GPG signature),
+**not** Maven Central. The unsuffixed `srcmorph-cli-<version>-jar-with-dependencies.jar` bundles the
+multi-platform **CPU** `net.ladenthin:llama` native and runs on any OS/arch. For GPU acceleration there
+is one classifier variant per llama native backend —
+`srcmorph-cli-<version>-jar-with-dependencies-<classifier>.jar` (e.g. `-cuda13-linux-x86-64`,
+`-vulkan-windows-x86-64`) — each bundling exactly that one backend's binary, so it runs only on the
+matching OS/arch with that GPU runtime present. Pick the plain CPU jar unless you specifically want a
+GPU build.
+
+Both `.json`/`.js` (parsed via a
 Jackson `ObjectMapper`) and `.yaml`/`.yml` (parsed via `YAMLMapper`) are supported — pick whichever you
 prefer; both mappers are configured strictly (`FAIL_ON_UNKNOWN_PROPERTIES`), so a typo'd key fails the
 run immediately rather than being silently ignored. On startup the CLI logs the parsed configuration
diff --git a/srcmorph-cli/pom.xml b/srcmorph-cli/pom.xml
index 1d442fe..c4495db 100644
--- a/srcmorph-cli/pom.xml
+++ b/srcmorph-cli/pom.xml
@@ -488,8 +488,18 @@ SPDX-License-Identifier: Apache-2.0
                 IS the deliverable artifact (the thing a user actually downloads and runs with
                 `java -jar`), not an occasional extra. Binding it in every `mvn package` (and therefore
                 every `mvn verify` / release build) keeps it in lockstep with the plain jar rather than
-                depending on a profile flag being remembered at release time. It rides `mvn -P release
-                deploy` as the `jar-with-dependencies` classifier alongside the plain jar.
+                depending on a profile flag being remembered at release time.
+
+                `<attach>false</attach>`: the fat jar is a GitHub-Release-only asset and must NOT be
+                deployed to Maven Central. A jar-with-dependencies on Central is redundant (consumers
+                depend on the plain `srcmorph-cli` jar + its resolved dependency graph, never the fat
+                jar) and, because it bundles the `net.ladenthin:llama` native binary, is both large and
+                platform-specific. Leaving it attached uploaded a 74 MB fat jar to Central on every
+                release. With attach=false the file is still produced in target/ (so the CI build's
+                upload-artifact and the release job's classifier-fat-jar loop still pick it up) but is
+                neither installed to ~/.m2 nor deployed to Central. Signing + GitHub attachment (for
+                this default CPU jar and one fat jar per net.ladenthin:llama GPU classifier) is done
+                explicitly by the publish-release / publish-snapshot jobs in .github/workflows/publish.yml.
             -->
             <plugin>
                 <groupId>org.apache.maven.plugins</groupId>
@@ -503,6 +513,7 @@ SPDX-License-Identifier: Apache-2.0
                             <mainClass>net.ladenthin.srcmorph.cli.Main</mainClass>
                         </manifest>
                     </archive>
+                    <attach>false</attach>
                 </configuration>
                 <executions>
                     <execution>