diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 72003ec2..b4ec2ad5 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -68,6 +68,56 @@ jobs:
     steps:
       - run: echo "Start gate elapsed — proceeding with pipeline."
 
+  # ---------------------------------------------------------------------------
+  # Download + cache the GGUF test models ONCE, upfront, for the whole pipeline.
+  # Every Java test job (`test-java-*`) and the langchain4j integration job `needs:` this
+  # job and then only RESTORES the shared cache (key gguf-models-v1) — so the ~5 GB model
+  # set is fetched from HuggingFace at most once per cache lifetime instead of racing to
+  # download in each job. GGUF is platform-independent, so this single ubuntu job's cache
+  # is reused by the macOS and Windows jobs too. On a warm cache this job is a no-op
+  # restore; on a cold cache it downloads all models, validates them, and saves the cache
+  # at job end (immutable key, so downstream jobs restore-hit). This is the single source
+  # of the download logic — do not re-add per-job downloads.
+  # ---------------------------------------------------------------------------
+
+  download-models:
+    name: Download + cache GGUF models (once)
+    needs: startgate
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+        uses: actions/cache@v6
+        with:
+          path: models/
+          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
+          # bump the suffix when the model set / URLs change.
+          key: gguf-models-v1
+      - name: Download text generation model
+        run: test -f models/${MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${MODEL_URL} --create-dirs -o models/${MODEL_NAME}
+      - name: Download reranking model
+        run: test -f models/${RERANKING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${RERANKING_MODEL_URL} --create-dirs -o models/${RERANKING_MODEL_NAME}
+      - name: Download draft model
+        run: test -f models/${DRAFT_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${DRAFT_MODEL_URL} --create-dirs -o models/${DRAFT_MODEL_NAME}
+      - name: Download reasoning model
+        run: test -f models/${REASONING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${REASONING_MODEL_URL} --create-dirs -o models/${REASONING_MODEL_NAME}
+      - name: Download tool-calling model
+        run: test -f models/${TOOL_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TOOL_MODEL_URL} --create-dirs -o models/${TOOL_MODEL_NAME}
+      - name: Download nomic embedding model (issue #98 regression)
+        run: test -f models/${NOMIC_EMBED_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${NOMIC_EMBED_MODEL_URL} --create-dirs -o models/${NOMIC_EMBED_MODEL_NAME}
+      - name: Download vision model
+        run: test -f models/${VISION_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MODEL_URL} --create-dirs -o models/${VISION_MODEL_NAME}
+      - name: Download vision mmproj
+        run: test -f models/${VISION_MMPROJ_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MMPROJ_URL} --create-dirs -o models/${VISION_MMPROJ_NAME}
+      - name: Download TTS model (OuteTTS)
+        run: test -f models/${TTS_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_MODEL_URL} --create-dirs -o models/${TTS_MODEL_NAME}
+      - name: Download TTS vocoder (WavTokenizer)
+        run: test -f models/${TTS_VOCODER_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_VOCODER_URL} --create-dirs -o models/${TTS_VOCODER_NAME}
+      - name: List files in models directory
+        run: ls -l models/
+      - name: Validate model files
+        run: bash .github/validate-models.sh
+
   # ---------------------------------------------------------------------------
   # Cross-compile jobs (Docker / dockcross) — produce release artifacts, no testing
   # ---------------------------------------------------------------------------
@@ -93,6 +143,87 @@ jobs:
           echo "=== internal package dependency graph (jdeps, bytecode) ==="
           jdeps -verbose:package target/classes | grep 'net.ladenthin.llama' || true
 
+  # ---------------------------------------------------------------------------
+  # Sibling module `llama-langchain4j` (LangChain4j adapters). Pure Java, no native
+  # code and no per-classifier matrix: it compiles against the core's stable Java API
+  # (identical across every classifier) and the backend is a runtime choice for the
+  # consumer. This job installs the core Java jar, guards that the module version is in
+  # lockstep with the core, then builds + tests the module (Java 17; langchain4j 1.x
+  # baseline). It runs its mapping unit tests; the model-backed integration test
+  # self-skips without a GGUF. `verify` also builds the javadoc/sources jars so a
+  # release-time javadoc break is caught here in PR CI.
+  # ---------------------------------------------------------------------------
+
+  test-java-llama-langchain4j:
+    name: Build and Test llama-langchain4j
+    needs: startgate
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/setup-java@v5
+        with:
+          java-version: ${{ env.JAVA_VERSION }}
+          distribution: temurin
+      - name: Version lockstep guard (module version must equal core version)
+        shell: bash
+        run: |
+          CORE=$(mvn -q -DforceStdout help:evaluate -Dexpression=project.version | tail -n1)
+          MOD=$(mvn -q -DforceStdout -f llama-langchain4j/pom.xml help:evaluate -Dexpression=project.version | tail -n1)
+          echo "core=$CORE module=$MOD"
+          if [ "$CORE" != "$MOD" ]; then
+            echo "::error::Version drift: core ($CORE) != llama-langchain4j ($MOD). Keep llama-langchain4j/pom.xml <version> in lockstep with the root pom.xml."
+            exit 1
+          fi
+      - name: Install core net.ladenthin:llama jar (Java only)
+        run: >
+          mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true
+          -Dspotless.check.skip=true -Dspotbugs.skip=true
+          -Dmaven.javadoc.skip=true -Dmaven.source.skip=true -Dgpg.skip=true install
+      - name: Build and test llama-langchain4j
+        run: mvn -B --no-transfer-progress -f llama-langchain4j/pom.xml verify
+
+  # ---------------------------------------------------------------------------
+  # Model-backed integration for the langchain4j adapters. Reuses the SAME shared GGUF
+  # cache (populated once by download-models) and the SAME Linux-x86_64 native artifact the
+  # core Java jobs already use — no extra model download and no duplicated download logic
+  # (restore-only cache, no curl steps). It exercises the chat / embedding / scoring
+  # adapters against the already-cached chat (Qwen3-0.6B), nomic-embedding and jina-reranker
+  # models. The model-backed tests self-skip when a model is absent, so a cold cache degrades
+  # to a skip, never a failure.
+  # ---------------------------------------------------------------------------
+
+  test-java-llama-langchain4j-integration:
+    name: Integration Test llama-langchain4j (model-backed)
+    needs: [crosscompile-linux-x86_64, download-models]
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - name: Download Linux x86_64 native library (reused, not rebuilt)
+        uses: actions/download-artifact@v8
+        with:
+          name: Linux-x86_64-libraries
+          path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
+        uses: actions/cache@v6
+        with:
+          path: models/
+          key: gguf-models-v1
+      - uses: actions/setup-java@v5
+        with:
+          distribution: 'temurin'
+          java-version: ${{ env.JAVA_VERSION }}
+      - name: Install core net.ladenthin:llama jar (bundles the downloaded native library)
+        run: >
+          mvn -B --no-transfer-progress -DskipTests -Denforcer.skip=true
+          -Dspotless.check.skip=true -Dspotbugs.skip=true
+          -Dmaven.javadoc.skip=true -Dmaven.source.skip=true -Dgpg.skip=true install
+      - name: Run llama-langchain4j model-backed integration tests (reused cached models)
+        run: >
+          mvn -B --no-transfer-progress -f llama-langchain4j/pom.xml test
+          -Dnet.ladenthin.llama.model.path=models/${REASONING_MODEL_NAME}
+          -Dnet.ladenthin.llama.langchain4j.embedding.model=models/${NOMIC_EMBED_MODEL_NAME}
+          -Dnet.ladenthin.llama.langchain4j.rerank.model=models/${RERANKING_MODEL_NAME}
+
   # ---------------------------------------------------------------------------
   # Build the llama.cpp WebUI ONCE, from the same pinned tag CMakeLists.txt fetches,
   # and share it to every native build as the generated, platform-independent
@@ -908,7 +1039,7 @@ jobs:
 
   test-java-linux-x86_64:
     name: Java Tests Ubuntu Latest x86_64
-    needs: crosscompile-linux-x86_64
+    needs: [crosscompile-linux-x86_64, download-models]
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
@@ -924,41 +1055,16 @@ jobs:
         with:
           name: Linux-x86_64-libraries
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      # GGUF model cache — introduced to stop re-downloading ~5 GB of test models from
-      # HuggingFace on every run (also dodges HF rate-limits). Complements the sccache compiler
-      # cache but is always ON: there is intentionally NO on/off flag for it (it is GitHub's
-      # free cache, safe + free), whereas the sccache cache is toggled by the `use_cache`
-      # workflow_dispatch input / USE_CACHE env. Not Depot — GB-scale blobs are usage-priced
-      # there and its file cache needs Depot-hosted runners. See CLAUDE.md.
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
+      # one entry (key gguf-models-v1). validate-models is kept as an integrity guard so a
+      # partial/absent restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: test -f models/${MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${MODEL_URL} --create-dirs -o models/${MODEL_NAME}
-      - name: Download reranking model
-        run: test -f models/${RERANKING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${RERANKING_MODEL_URL} --create-dirs -o models/${RERANKING_MODEL_NAME}
-      - name: Download draft model
-        run: test -f models/${DRAFT_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${DRAFT_MODEL_URL} --create-dirs -o models/${DRAFT_MODEL_NAME}
-      - name: Download reasoning model
-        run: test -f models/${REASONING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${REASONING_MODEL_URL} --create-dirs -o models/${REASONING_MODEL_NAME}
-      - name: Download tool-calling model
-        run: test -f models/${TOOL_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TOOL_MODEL_URL} --create-dirs -o models/${TOOL_MODEL_NAME}
-      - name: Download nomic embedding model (issue #98 regression)
-        run: test -f models/${NOMIC_EMBED_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${NOMIC_EMBED_MODEL_URL} --create-dirs -o models/${NOMIC_EMBED_MODEL_NAME}
-      - name: Download vision model
-        run: test -f models/${VISION_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MODEL_URL} --create-dirs -o models/${VISION_MODEL_NAME}
-      - name: Download vision mmproj
-        run: test -f models/${VISION_MMPROJ_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MMPROJ_URL} --create-dirs -o models/${VISION_MMPROJ_NAME}
-      - name: Download TTS model (OuteTTS)
-        run: test -f models/${TTS_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_MODEL_URL} --create-dirs -o models/${TTS_MODEL_NAME}
-      - name: Download TTS vocoder (WavTokenizer)
-        run: test -f models/${TTS_VOCODER_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_VOCODER_URL} --create-dirs -o models/${TTS_VOCODER_NAME}
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: bash .github/validate-models.sh
       - uses: actions/setup-java@v5
@@ -1051,7 +1157,7 @@ jobs:
 
   test-java-macos-arm64-metal:
     name: Java Tests macOS 14 arm64 (Metal)
-    needs: build-macos-arm64-metal
+    needs: [build-macos-arm64-metal, download-models]
     runs-on: macos-14
     steps:
       - uses: actions/checkout@v7
@@ -1067,41 +1173,16 @@ jobs:
         with:
           name: macos-14-libraries
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      # GGUF model cache — introduced to stop re-downloading ~5 GB of test models from
-      # HuggingFace on every run (also dodges HF rate-limits). Complements the sccache compiler
-      # cache but is always ON: there is intentionally NO on/off flag for it (it is GitHub's
-      # free cache, safe + free), whereas the sccache cache is toggled by the `use_cache`
-      # workflow_dispatch input / USE_CACHE env. Not Depot — GB-scale blobs are usage-priced
-      # there and its file cache needs Depot-hosted runners. See CLAUDE.md.
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
+      # one entry (key gguf-models-v1). validate-models is kept as an integrity guard so a
+      # partial/absent restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: test -f models/${MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${MODEL_URL} --create-dirs -o models/${MODEL_NAME}
-      - name: Download reranking model
-        run: test -f models/${RERANKING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${RERANKING_MODEL_URL} --create-dirs -o models/${RERANKING_MODEL_NAME}
-      - name: Download draft model
-        run: test -f models/${DRAFT_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${DRAFT_MODEL_URL} --create-dirs -o models/${DRAFT_MODEL_NAME}
-      - name: Download reasoning model
-        run: test -f models/${REASONING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${REASONING_MODEL_URL} --create-dirs -o models/${REASONING_MODEL_NAME}
-      - name: Download tool-calling model
-        run: test -f models/${TOOL_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TOOL_MODEL_URL} --create-dirs -o models/${TOOL_MODEL_NAME}
-      - name: Download nomic embedding model (issue #98 regression)
-        run: test -f models/${NOMIC_EMBED_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${NOMIC_EMBED_MODEL_URL} --create-dirs -o models/${NOMIC_EMBED_MODEL_NAME}
-      - name: Download vision model
-        run: test -f models/${VISION_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MODEL_URL} --create-dirs -o models/${VISION_MODEL_NAME}
-      - name: Download vision mmproj
-        run: test -f models/${VISION_MMPROJ_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MMPROJ_URL} --create-dirs -o models/${VISION_MMPROJ_NAME}
-      - name: Download TTS model (OuteTTS)
-        run: test -f models/${TTS_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_MODEL_URL} --create-dirs -o models/${TTS_MODEL_NAME}
-      - name: Download TTS vocoder (WavTokenizer)
-        run: test -f models/${TTS_VOCODER_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_VOCODER_URL} --create-dirs -o models/${TTS_VOCODER_NAME}
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: bash .github/validate-models.sh
       - uses: actions/setup-java@v5
@@ -1140,7 +1221,7 @@ jobs:
 
   test-java-macos-arm64-no-metal:
     name: Java Tests macOS 15 arm64 (no Metal)
-    needs: build-macos-arm64-no-metal
+    needs: [build-macos-arm64-no-metal, download-models]
     runs-on: macos-15
     steps:
       - uses: actions/checkout@v7
@@ -1156,41 +1237,16 @@ jobs:
         with:
           name: macos-15-libraries
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      # GGUF model cache — introduced to stop re-downloading ~5 GB of test models from
-      # HuggingFace on every run (also dodges HF rate-limits). Complements the sccache compiler
-      # cache but is always ON: there is intentionally NO on/off flag for it (it is GitHub's
-      # free cache, safe + free), whereas the sccache cache is toggled by the `use_cache`
-      # workflow_dispatch input / USE_CACHE env. Not Depot — GB-scale blobs are usage-priced
-      # there and its file cache needs Depot-hosted runners. See CLAUDE.md.
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
+      # one entry (key gguf-models-v1). validate-models is kept as an integrity guard so a
+      # partial/absent restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: test -f models/${MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${MODEL_URL} --create-dirs -o models/${MODEL_NAME}
-      - name: Download reranking model
-        run: test -f models/${RERANKING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${RERANKING_MODEL_URL} --create-dirs -o models/${RERANKING_MODEL_NAME}
-      - name: Download draft model
-        run: test -f models/${DRAFT_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${DRAFT_MODEL_URL} --create-dirs -o models/${DRAFT_MODEL_NAME}
-      - name: Download reasoning model
-        run: test -f models/${REASONING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${REASONING_MODEL_URL} --create-dirs -o models/${REASONING_MODEL_NAME}
-      - name: Download tool-calling model
-        run: test -f models/${TOOL_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TOOL_MODEL_URL} --create-dirs -o models/${TOOL_MODEL_NAME}
-      - name: Download nomic embedding model (issue #98 regression)
-        run: test -f models/${NOMIC_EMBED_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${NOMIC_EMBED_MODEL_URL} --create-dirs -o models/${NOMIC_EMBED_MODEL_NAME}
-      - name: Download vision model
-        run: test -f models/${VISION_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MODEL_URL} --create-dirs -o models/${VISION_MODEL_NAME}
-      - name: Download vision mmproj
-        run: test -f models/${VISION_MMPROJ_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MMPROJ_URL} --create-dirs -o models/${VISION_MMPROJ_NAME}
-      - name: Download TTS model (OuteTTS)
-        run: test -f models/${TTS_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_MODEL_URL} --create-dirs -o models/${TTS_MODEL_NAME}
-      - name: Download TTS vocoder (WavTokenizer)
-        run: test -f models/${TTS_VOCODER_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_VOCODER_URL} --create-dirs -o models/${TTS_VOCODER_NAME}
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: bash .github/validate-models.sh
       - uses: actions/setup-java@v5
@@ -1229,7 +1285,7 @@ jobs:
 
   test-java-macos-arm64-metal-15:
     name: Java Tests macOS 15 arm64 (Metal)
-    needs: build-macos-arm64-metal-15
+    needs: [build-macos-arm64-metal-15, download-models]
     runs-on: macos-15
     steps:
       - uses: actions/checkout@v7
@@ -1245,41 +1301,16 @@ jobs:
         with:
           name: macos-15-metal-libraries
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      # GGUF model cache — introduced to stop re-downloading ~5 GB of test models from
-      # HuggingFace on every run (also dodges HF rate-limits). Complements the sccache compiler
-      # cache but is always ON: there is intentionally NO on/off flag for it (it is GitHub's
-      # free cache, safe + free), whereas the sccache cache is toggled by the `use_cache`
-      # workflow_dispatch input / USE_CACHE env. Not Depot — GB-scale blobs are usage-priced
-      # there and its file cache needs Depot-hosted runners. See CLAUDE.md.
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. GGUF is platform-independent, so ubuntu + macOS + Windows share
+      # one entry (key gguf-models-v1). validate-models is kept as an integrity guard so a
+      # partial/absent restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: test -f models/${MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${MODEL_URL} --create-dirs -o models/${MODEL_NAME}
-      - name: Download reranking model
-        run: test -f models/${RERANKING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${RERANKING_MODEL_URL} --create-dirs -o models/${RERANKING_MODEL_NAME}
-      - name: Download draft model
-        run: test -f models/${DRAFT_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${DRAFT_MODEL_URL} --create-dirs -o models/${DRAFT_MODEL_NAME}
-      - name: Download reasoning model
-        run: test -f models/${REASONING_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${REASONING_MODEL_URL} --create-dirs -o models/${REASONING_MODEL_NAME}
-      - name: Download tool-calling model
-        run: test -f models/${TOOL_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TOOL_MODEL_URL} --create-dirs -o models/${TOOL_MODEL_NAME}
-      - name: Download nomic embedding model (issue #98 regression)
-        run: test -f models/${NOMIC_EMBED_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${NOMIC_EMBED_MODEL_URL} --create-dirs -o models/${NOMIC_EMBED_MODEL_NAME}
-      - name: Download vision model
-        run: test -f models/${VISION_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MODEL_URL} --create-dirs -o models/${VISION_MODEL_NAME}
-      - name: Download vision mmproj
-        run: test -f models/${VISION_MMPROJ_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${VISION_MMPROJ_URL} --create-dirs -o models/${VISION_MMPROJ_NAME}
-      - name: Download TTS model (OuteTTS)
-        run: test -f models/${TTS_MODEL_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_MODEL_URL} --create-dirs -o models/${TTS_MODEL_NAME}
-      - name: Download TTS vocoder (WavTokenizer)
-        run: test -f models/${TTS_VOCODER_NAME} || curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors ${TTS_VOCODER_URL} --create-dirs -o models/${TTS_VOCODER_NAME}
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: bash .github/validate-models.sh
       - uses: actions/setup-java@v5
@@ -1318,7 +1349,7 @@ jobs:
 
   test-java-windows-x86_64:
     name: Java Tests Windows 2025 x86_64 (default / Ninja)
-    needs: build-windows-x86_64
+    needs: [build-windows-x86_64, download-models]
     runs-on: windows-2025-vs2026
     steps:
       - uses: actions/checkout@v7
@@ -1337,35 +1368,15 @@ jobs:
         with:
           name: Windows-x86_64-libraries
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. validate-models is kept as an integrity guard so a partial/absent
+      # restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: if (-not (Test-Path "models/$env:MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:MODEL_URL --create-dirs -o models/$env:MODEL_NAME }
-      - name: Download reranking model
-        run: if (-not (Test-Path "models/$env:RERANKING_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:RERANKING_MODEL_URL --create-dirs -o models/$env:RERANKING_MODEL_NAME }
-      - name: Download draft model
-        run: if (-not (Test-Path "models/$env:DRAFT_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:DRAFT_MODEL_URL --create-dirs -o models/$env:DRAFT_MODEL_NAME }
-      - name: Download reasoning model
-        run: if (-not (Test-Path "models/$env:REASONING_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:REASONING_MODEL_URL --create-dirs -o models/$env:REASONING_MODEL_NAME }
-      - name: Download tool-calling model
-        run: if (-not (Test-Path "models/$env:TOOL_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TOOL_MODEL_URL --create-dirs -o models/$env:TOOL_MODEL_NAME }
-      - name: Download nomic embedding model (issue #98 regression)
-        run: if (-not (Test-Path "models/$env:NOMIC_EMBED_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:NOMIC_EMBED_MODEL_URL --create-dirs -o models/$env:NOMIC_EMBED_MODEL_NAME }
-      - name: Download vision model
-        run: if (-not (Test-Path "models/$env:VISION_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:VISION_MODEL_URL --create-dirs -o models/$env:VISION_MODEL_NAME }
-      - name: Download vision mmproj
-        run: if (-not (Test-Path "models/$env:VISION_MMPROJ_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:VISION_MMPROJ_URL --create-dirs -o models/$env:VISION_MMPROJ_NAME }
-      - name: Download TTS model (OuteTTS)
-        run: if (-not (Test-Path "models/$env:TTS_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TTS_MODEL_URL --create-dirs -o models/$env:TTS_MODEL_NAME }
-      - name: Download TTS vocoder (WavTokenizer)
-        run: if (-not (Test-Path "models/$env:TTS_VOCODER_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TTS_VOCODER_URL --create-dirs -o models/$env:TTS_VOCODER_NAME }
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: .github\validate-models.bat
       - uses: actions/setup-java@v5
@@ -1427,7 +1438,7 @@ jobs:
   # validated end-to-end before the `msvc-windows` classifier JAR ships.
   test-java-windows-x86_64-msvc:
     name: Java Tests Windows 2025 x86_64 (MSVC classifier)
-    needs: build-windows-x86_64-msvc
+    needs: [build-windows-x86_64-msvc, download-models]
     runs-on: windows-2025-vs2026
     steps:
       - uses: actions/checkout@v7
@@ -1446,35 +1457,15 @@ jobs:
         with:
           name: Windows-x86_64-msvc
           path: ${{ github.workspace }}/src/main/resources/net/ladenthin/llama/
-      - name: Cache GGUF models (GitHub Actions cache; avoids re-downloading from HuggingFace)
+      # GGUF models are downloaded + cached ONCE by the upstream `download-models` job
+      # (this job `needs:` it), so here we only RESTORE the shared cache — no per-job
+      # download logic. validate-models is kept as an integrity guard so a partial/absent
+      # restore fails loudly instead of silently self-skipping required tests.
+      - name: Restore shared GGUF model cache (populated by download-models; no re-download)
         uses: actions/cache@v6
         with:
           path: models/
-          # GGUF is platform-independent, so ubuntu + macOS + Windows share one entry;
-          # bump the suffix when the model set / URLs change.
           key: gguf-models-v1
-      - name: Download text generation model
-        run: if (-not (Test-Path "models/$env:MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:MODEL_URL --create-dirs -o models/$env:MODEL_NAME }
-      - name: Download reranking model
-        run: if (-not (Test-Path "models/$env:RERANKING_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:RERANKING_MODEL_URL --create-dirs -o models/$env:RERANKING_MODEL_NAME }
-      - name: Download draft model
-        run: if (-not (Test-Path "models/$env:DRAFT_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:DRAFT_MODEL_URL --create-dirs -o models/$env:DRAFT_MODEL_NAME }
-      - name: Download reasoning model
-        run: if (-not (Test-Path "models/$env:REASONING_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:REASONING_MODEL_URL --create-dirs -o models/$env:REASONING_MODEL_NAME }
-      - name: Download tool-calling model
-        run: if (-not (Test-Path "models/$env:TOOL_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TOOL_MODEL_URL --create-dirs -o models/$env:TOOL_MODEL_NAME }
-      - name: Download nomic embedding model (issue #98 regression)
-        run: if (-not (Test-Path "models/$env:NOMIC_EMBED_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:NOMIC_EMBED_MODEL_URL --create-dirs -o models/$env:NOMIC_EMBED_MODEL_NAME }
-      - name: Download vision model
-        run: if (-not (Test-Path "models/$env:VISION_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:VISION_MODEL_URL --create-dirs -o models/$env:VISION_MODEL_NAME }
-      - name: Download vision mmproj
-        run: if (-not (Test-Path "models/$env:VISION_MMPROJ_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:VISION_MMPROJ_URL --create-dirs -o models/$env:VISION_MMPROJ_NAME }
-      - name: Download TTS model (OuteTTS)
-        run: if (-not (Test-Path "models/$env:TTS_MODEL_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TTS_MODEL_URL --create-dirs -o models/$env:TTS_MODEL_NAME }
-      - name: Download TTS vocoder (WavTokenizer)
-        run: if (-not (Test-Path "models/$env:TTS_VOCODER_NAME")) { curl -L --proto =https --proto-redir =https --fail --retry 5 --retry-all-errors $env:TTS_VOCODER_URL --create-dirs -o models/$env:TTS_VOCODER_NAME }
-      - name: List files in models directory
-        run: ls -l models/
       - name: Validate model files
         run: .github\validate-models.bat
       - uses: actions/setup-java@v5
@@ -1665,7 +1656,7 @@ jobs:
 
   publish-snapshot:
     name: Publish Snapshot to Central
-    needs: [check-snapshot, crosscompile-linux-x86_64-cuda, crosscompile-android-aarch64-opencl, code-style]
+    needs: [check-snapshot, crosscompile-linux-x86_64-cuda, crosscompile-android-aarch64-opencl, code-style, test-java-llama-langchain4j]
     if: needs.check-snapshot.result == 'success' && inputs.publish_to_central
     runs-on: ubuntu-latest
     environment: maven-central
@@ -1731,11 +1722,23 @@ jobs:
           MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
           MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
           MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+      # Deploy the sibling llama-langchain4j at the same version. The core `deploy`
+      # above ran through `install`, so net.ladenthin:llama is in the local repo for
+      # the module to resolve. Standalone module (not in the reactor), so it is a
+      # separate deploy invocation.
+      - name: Publish snapshot (llama-langchain4j)
+        run: mvn --batch-mode --no-transfer-progress -f llama-langchain4j/pom.xml -P release -Dmaven.test.skip=true deploy
+        env:
+          MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
+          MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
+          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
       - name: Collect signed artifacts
         run: |
           mkdir -p signed-snapshot-assets
           cp target/*.jar signed-snapshot-assets/ 2>/dev/null || true
           cp target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
+          cp llama-langchain4j/target/*.jar signed-snapshot-assets/ 2>/dev/null || true
+          cp llama-langchain4j/target/*.jar.asc signed-snapshot-assets/ 2>/dev/null || true
       - uses: actions/upload-artifact@v7
         with:
           name: signed-snapshot-assets
@@ -1770,7 +1773,7 @@ jobs:
   publish-release:
     name: Publish Release to Central
     if: needs.check-tag.result == 'success' && inputs.publish_to_central
-    needs: [check-tag, crosscompile-linux-x86_64-cuda, crosscompile-android-aarch64-opencl, code-style]
+    needs: [check-tag, crosscompile-linux-x86_64-cuda, crosscompile-android-aarch64-opencl, code-style, test-java-llama-langchain4j]
     runs-on: ubuntu-latest
     environment: maven-central
     permissions:
@@ -1826,11 +1829,23 @@ jobs:
           MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
           MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
           MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+      # Deploy the sibling llama-langchain4j at the same version. The core `deploy`
+      # above ran through `install`, so net.ladenthin:llama is in the local repo for
+      # the module to resolve. Standalone module (not in the reactor), so it is a
+      # separate deploy invocation.
+      - name: Publish release (llama-langchain4j)
+        run: mvn --batch-mode --no-transfer-progress -f llama-langchain4j/pom.xml -P release -Dmaven.test.skip=true deploy
+        env:
+          MAVEN_USERNAME: ${{ secrets.CENTRAL_USERNAME }}
+          MAVEN_PASSWORD: ${{ secrets.CENTRAL_TOKEN }}
+          MAVEN_GPG_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
       - name: Collect signed artifacts
         run: |
           mkdir -p signed-release-assets
           cp target/*.jar signed-release-assets/ 2>/dev/null || true
           cp target/*.jar.asc signed-release-assets/ 2>/dev/null || true
+          cp llama-langchain4j/target/*.jar signed-release-assets/ 2>/dev/null || true
+          cp llama-langchain4j/target/*.jar.asc signed-release-assets/ 2>/dev/null || true
       - uses: actions/upload-artifact@v7
         with:
           name: signed-release-assets
diff --git a/CLAUDE.md b/CLAUDE.md
index 7139beef..db03010f 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -971,12 +971,18 @@ properties set, so `LlamaEmbeddingsTest`, `MultimodalIntegrationTest`, and `TtsI
 these as **required** (a missing model hard-fails the job before tests run, so a download
 regression can never silently downgrade to a skip). The only model still self-skipping is the
 audio-input model (`AudioInputIntegrationTest`) — it has no committed clip and no CI download.
-The shared GGUF cache (`actions/cache`, key `gguf-models-v1`, path `models/`) holds the full set;
-since every test job downloads the full set before the cache can save, whichever job wins the
-save race caches everything. Because the cache key is immutable, changing the model set means the
-**existing cache entry must be deleted** (not bumped to `v2`) so the next run rebuilds it complete
-— locally the model tests still self-skip when a GGUF is absent (`Assume.assumeTrue`), so a
-partial local checkout is fine.
+The shared GGUF cache (`actions/cache`, key `gguf-models-v1`, path `models/`) holds the full set
+and is populated **once, upfront** by a dedicated **`download-models`** job (`needs: startgate`):
+it is the single place the ~5 GB set is fetched from HuggingFace (the ten `curl` steps + the
+`validate-models.sh` gate live only there). Every `test-java-*` job — and the langchain4j
+integration job — `needs: download-models` and then only **restores** that cache (no per-job
+download, no cold-start save race), keeping `validate-models.{sh,bat}` as a per-job integrity
+guard. GGUF is platform-independent, so the one ubuntu `download-models` cache is reused by the
+macOS and Windows jobs too. `validate-models.{sh,bat}` treats the models as **required** (a
+missing model hard-fails the job before tests run). Because the cache key is immutable, changing
+the model set means the **existing cache entry must be deleted** (not bumped to `v2`) so
+`download-models` rebuilds it complete — locally the model tests still self-skip when a GGUF is
+absent (`Assume.assumeTrue`), so a partial local checkout is fine.
 
 Set the model path via system property or environment variable (see test files for exact property names).
 
@@ -1219,6 +1225,56 @@ keeping it clear of the JPMS module-mode javadoc trap that bit BAF. **Before rai
 javadoc source level to ≥ 9, read**
 [`../workspace/policies/jpms-module-descriptor.md`](../workspace/policies/jpms-module-descriptor.md).
 
+## LangChain4j integration (`llama-langchain4j` sibling module)
+
+`llama-langchain4j/` adapts a `LlamaModel` to LangChain4j's `ChatModel`,
+`StreamingChatModel`, `EmbeddingModel` and `ScoringModel` interfaces **in-process over
+JNI** (no HTTP hop). It is a **standalone sibling module**, deliberately *not* in the root
+reactor, so the native build/release pipeline is untouched.
+
+Why it is a **separate artifact** and not a classifier of the core: langchain4j 1.x
+requires **Java 17** (the core stays Java 8), and classifiers share the core's single POM —
+adding `langchain4j-core` there would force it (and the Java 17 floor) on every plain
+`net.ladenthin:llama` consumer. A separate `artifactId` with its own POM is the only way to
+keep that dependency (and Java floor) off the core. It is pure Java with **no per-classifier
+matrix**: it compiles against the core's Java API, which is identical across every native
+classifier; the backend (CPU/CUDA/OpenCL/Vulkan) is a runtime classpath choice for the
+consumer.
+
+Wiring:
+
+1. **`llama-langchain4j/pom.xml`** — `net.ladenthin:llama-langchain4j`, `release 17`,
+   depends on `net.ladenthin:llama:${project.version}` (so the core dep always matches the
+   module's own version) and `dev.langchain4j:langchain4j-core`. Carries its own
+   sources/javadoc/gpg + `release` profile (Central requires per-artifact signing; the module
+   has no parent to inherit them from — plugin versions are pinned in lockstep with the root
+   `pom.xml`). Java package stays `net.ladenthin.llama.langchain4j` (package name need not track
+   the artifactId).
+2. **`.github/workflows/publish.yml`** — the `test-java-llama-langchain4j` job installs the
+   core Java jar, runs a **version-lockstep guard** (module version must equal core version,
+   else the build fails — the standalone module can't inherit `${project.version}` from a
+   reactor), then `mvn -f llama-langchain4j/pom.xml verify` (7 model-free mapping unit tests
+   run; the 4 model-backed integration tests self-skip without a GGUF; `verify` also builds the
+   javadoc jar so a release-time javadoc break is caught in PR CI). The
+   `publish-snapshot`/`publish-release` jobs `needs:` this job and, after the core `deploy`
+   (which installs the core jar locally), run a second `deploy` of the module at the same
+   version. A separate **`test-java-llama-langchain4j-integration`** job runs the model-backed
+   tests (chat/streaming/embedding/scoring adapters) by **reusing** the shared GGUF cache
+   (`gguf-models-v1`, restore-only — no extra download) and the `Linux-x86_64-libraries` native
+   artifact: it `needs: [crosscompile-linux-x86_64, download-models]` (so the cache is already
+   populated and it runs in parallel), installs the core jar with the downloaded native lib
+   bundled, and passes the already-cached chat
+   (`REASONING_MODEL_NAME`), nomic-embedding and jina-reranker model paths via the module's
+   `-Dnet.ladenthin.llama.langchain4j.{embedding,rerank}.model` / `net.ladenthin.llama.model.path`
+   properties. It is validation-only (not a release gate); a cold cache degrades to a self-skip.
+3. **Version bumps** — when the root `pom.xml` `<version>` changes, bump
+   `llama-langchain4j/pom.xml` `<version>` to match in the same commit, or the lockstep guard
+   reds CI.
+
+**Open follow-ups** (documented in `llama-langchain4j/README.md`): tool calling
+(`ToolSpecification` ↔ jllama `ToolDefinition`), `response_format`/JSON mode, and multimodal
+user input (currently flattened to text).
+
 ## Open TODOs
 
 Open TODOs for this repo live in [`TODO.md`](TODO.md). Cross-repo status
diff --git a/REUSE.toml b/REUSE.toml
index 54d18d08..df4c5422 100644
--- a/REUSE.toml
+++ b/REUSE.toml
@@ -24,7 +24,7 @@ path = [
     ".github/ISSUE_TEMPLATE/bug_report.md",
     ".github/ISSUE_TEMPLATE/feature_request.md",
     ".claude/commands/find-cpp-duplication.md",
-    "langchain4j-jllama/README.md",
+    "llama-langchain4j/README.md",
 ]
 SPDX-FileCopyrightText = [
     "2023-2025 Konstantin Herud",
diff --git a/langchain4j-jllama/pom.xml b/langchain4j-jllama/pom.xml
deleted file mode 100644
index 58f6d365..00000000
--- a/langchain4j-jllama/pom.xml
+++ /dev/null
@@ -1,94 +0,0 @@
-<!--
-SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
-
-SPDX-License-Identifier: MIT
--->
-
-<project xmlns="http://maven.apache.org/POM/4.0.0"
-	xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
-	xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
-	<modelVersion>4.0.0</modelVersion>
-
-	<groupId>net.ladenthin</groupId>
-	<artifactId>langchain4j-jllama</artifactId>
-	<version>5.0.4-SNAPSHOT</version>
-	<packaging>jar</packaging>
-
-	<name>${project.groupId}:${project.artifactId}</name>
-	<description>LangChain4j integration for java-llama.cpp: in-process ChatModel,
-		StreamingChatModel, EmbeddingModel and ScoringModel adapters backed by a
-		llama.cpp model over JNI (no HTTP hop).</description>
-	<url>https://github.com/bernardladenthin/java-llama.cpp</url>
-
-	<licenses>
-		<license>
-			<name>MIT License</name>
-			<url>https://www.opensource.org/licenses/mit-license.php</url>
-			<distribution>repo</distribution>
-		</license>
-	</licenses>
-
-	<developers>
-		<developer>
-			<name>Bernard Ladenthin</name>
-			<organizationUrl>https://github.com/bernardladenthin</organizationUrl>
-		</developer>
-	</developers>
-
-	<scm>
-		<connection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</connection>
-		<developerConnection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</developerConnection>
-		<url>https://github.com/bernardladenthin/java-llama.cpp/tree/main</url>
-	</scm>
-
-	<properties>
-		<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
-		<maven.compiler.release>17</maven.compiler.release>
-		<!-- Keep in lockstep with the core java-llama.cpp artifact version. -->
-		<jllama.version>5.0.4-SNAPSHOT</jllama.version>
-		<langchain4j.version>1.17.1</langchain4j.version>
-		<junit.version>6.1.1</junit.version>
-		<hamcrest.version>3.0</hamcrest.version>
-		<surefire.version>3.5.5</surefire.version>
-	</properties>
-
-	<dependencies>
-		<!-- The JNI binding we adapt. Provided-by-the-consumer in spirit, but compile
-		     scope so a consumer that only declares langchain4j-jllama still gets it. -->
-		<dependency>
-			<groupId>net.ladenthin</groupId>
-			<artifactId>llama</artifactId>
-			<version>${jllama.version}</version>
-		</dependency>
-
-		<!-- The interfaces we implement (ChatModel/StreamingChatModel/EmbeddingModel/ScoringModel). -->
-		<dependency>
-			<groupId>dev.langchain4j</groupId>
-			<artifactId>langchain4j-core</artifactId>
-			<version>${langchain4j.version}</version>
-		</dependency>
-
-		<dependency>
-			<groupId>org.junit.jupiter</groupId>
-			<artifactId>junit-jupiter</artifactId>
-			<version>${junit.version}</version>
-			<scope>test</scope>
-		</dependency>
-		<dependency>
-			<groupId>org.hamcrest</groupId>
-			<artifactId>hamcrest</artifactId>
-			<version>${hamcrest.version}</version>
-			<scope>test</scope>
-		</dependency>
-	</dependencies>
-
-	<build>
-		<plugins>
-			<plugin>
-				<groupId>org.apache.maven.plugins</groupId>
-				<artifactId>maven-surefire-plugin</artifactId>
-				<version>${surefire.version}</version>
-			</plugin>
-		</plugins>
-	</build>
-</project>
diff --git a/langchain4j-jllama/README.md b/llama-langchain4j/README.md
similarity index 80%
rename from langchain4j-jllama/README.md
rename to llama-langchain4j/README.md
index cb9d99bd..1082d973 100644
--- a/langchain4j-jllama/README.md
+++ b/llama-langchain4j/README.md
@@ -1,4 +1,4 @@
-# langchain4j-jllama
+# llama-langchain4j
 
 [LangChain4j](https://github.com/langchain4j/langchain4j) adapters backed by an **in-process**
 [java-llama.cpp](https://github.com/bernardladenthin/java-llama.cpp) model over JNI — no HTTP server,
@@ -61,7 +61,7 @@ ScoringModel reranker     = new JllamaScoringModel(rerankLlama);
 ```xml
 <dependency>
     <groupId>net.ladenthin</groupId>
-    <artifactId>langchain4j-jllama</artifactId>
+    <artifactId>llama-langchain4j</artifactId>
     <version>5.0.4-SNAPSHOT</version>
 </dependency>
 ```
@@ -79,16 +79,28 @@ build here:
 mvn -DskipTests install
 
 # then build/test this module
-cd langchain4j-jllama
+cd llama-langchain4j
 mvn test
 ```
 
-The end-to-end test (`JllamaChatModelIntegrationTest`) self-skips unless you pass a model:
+The model-backed integration tests self-skip unless you point them at a GGUF. Each adapter has
+its own property so you can run them independently (a chat/instruct model, an embedding-mode model,
+and a reranking-mode model respectively):
 
 ```bash
-mvn test -Dnet.ladenthin.llama.model.path=/abs/path/to/model.gguf
+# chat + streaming (JllamaChatModelIntegrationTest)
+mvn test -Dnet.ladenthin.llama.model.path=/abs/path/to/chat.gguf
+# embeddings (JllamaEmbeddingModelIntegrationTest)
+mvn test -Dnet.ladenthin.llama.langchain4j.embedding.model=/abs/path/to/embedding.gguf
+# re-ranking / scoring (JllamaScoringModelIntegrationTest)
+mvn test -Dnet.ladenthin.llama.langchain4j.rerank.model=/abs/path/to/reranker.gguf
 ```
 
+In CI these reuse the project's existing shared GGUF cache (the chat, nomic-embedding and
+jina-reranker models the core test jobs already download) — the
+`test-java-llama-langchain4j-integration` job restores that cache and the
+`Linux-x86_64` native library artifact, so no extra model is downloaded.
+
 ## Not mapped yet
 
 - **Tool calling.** `ChatRequest.toolSpecifications()` are not forwarded, so the chat adapters return
diff --git a/llama-langchain4j/pom.xml b/llama-langchain4j/pom.xml
new file mode 100644
index 00000000..ecb0b8bb
--- /dev/null
+++ b/llama-langchain4j/pom.xml
@@ -0,0 +1,193 @@
+<!--
+SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
+
+SPDX-License-Identifier: MIT
+-->
+
+<project xmlns="http://maven.apache.org/POM/4.0.0"
+	xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
+	xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
+	<modelVersion>4.0.0</modelVersion>
+
+	<groupId>net.ladenthin</groupId>
+	<artifactId>llama-langchain4j</artifactId>
+	<version>5.0.4-SNAPSHOT</version>
+	<packaging>jar</packaging>
+
+	<name>${project.groupId}:${project.artifactId}</name>
+	<description>LangChain4j integration for java-llama.cpp: in-process ChatModel,
+		StreamingChatModel, EmbeddingModel and ScoringModel adapters backed by a
+		llama.cpp model over JNI (no HTTP hop).</description>
+	<url>https://github.com/bernardladenthin/java-llama.cpp</url>
+
+	<licenses>
+		<license>
+			<name>MIT License</name>
+			<url>https://www.opensource.org/licenses/mit-license.php</url>
+			<distribution>repo</distribution>
+		</license>
+	</licenses>
+
+	<developers>
+		<developer>
+			<name>Bernard Ladenthin</name>
+			<organizationUrl>https://github.com/bernardladenthin</organizationUrl>
+		</developer>
+	</developers>
+
+	<scm>
+		<connection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</connection>
+		<developerConnection>scm:git:https://github.com/bernardladenthin/java-llama.cpp.git</developerConnection>
+		<url>https://github.com/bernardladenthin/java-llama.cpp/tree/main</url>
+	</scm>
+
+	<distributionManagement>
+		<snapshotRepository>
+			<id>central</id>
+			<url>https://central.sonatype.com/repository/maven-snapshots/</url>
+		</snapshotRepository>
+	</distributionManagement>
+
+	<properties>
+		<project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
+		<!-- langchain4j 1.x requires Java 17; the core net.ladenthin:llama stays Java 8. -->
+		<maven.compiler.release>17</maven.compiler.release>
+		<langchain4j.version>1.17.1</langchain4j.version>
+		<junit.version>6.1.1</junit.version>
+		<hamcrest.version>3.0</hamcrest.version>
+		<!-- Plugin versions are kept in lockstep with the core pom.xml. This module has no
+		     parent, so it cannot inherit their pluginManagement and must pin them here. -->
+		<compiler.plugin.version>3.15.0</compiler.plugin.version>
+		<surefire.version>3.5.6</surefire.version>
+		<source.plugin.version>3.4.0</source.plugin.version>
+		<javadoc.plugin.version>3.12.0</javadoc.plugin.version>
+		<gpg.plugin.version>3.2.8</gpg.plugin.version>
+		<central.plugin.version>0.11.0</central.plugin.version>
+	</properties>
+
+	<dependencies>
+		<!-- The JNI binding we adapt. Version is pinned to this module's own version via
+		     ${project.version}: the module always ships in lockstep with the core artifact
+		     (a CI guard fails the build if the two versions ever diverge). -->
+		<dependency>
+			<groupId>net.ladenthin</groupId>
+			<artifactId>llama</artifactId>
+			<version>${project.version}</version>
+		</dependency>
+
+		<!-- The interfaces we implement (ChatModel/StreamingChatModel/EmbeddingModel/ScoringModel). -->
+		<dependency>
+			<groupId>dev.langchain4j</groupId>
+			<artifactId>langchain4j-core</artifactId>
+			<version>${langchain4j.version}</version>
+		</dependency>
+
+		<dependency>
+			<groupId>org.junit.jupiter</groupId>
+			<artifactId>junit-jupiter</artifactId>
+			<version>${junit.version}</version>
+			<scope>test</scope>
+		</dependency>
+		<dependency>
+			<groupId>org.hamcrest</groupId>
+			<artifactId>hamcrest</artifactId>
+			<version>${hamcrest.version}</version>
+			<scope>test</scope>
+		</dependency>
+	</dependencies>
+
+	<build>
+		<plugins>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-compiler-plugin</artifactId>
+				<version>${compiler.plugin.version}</version>
+			</plugin>
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-surefire-plugin</artifactId>
+				<version>${surefire.version}</version>
+			</plugin>
+
+			<!-- -sources.jar (Maven Central requirement). -->
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-source-plugin</artifactId>
+				<version>${source.plugin.version}</version>
+				<executions>
+					<execution>
+						<id>attach-sources</id>
+						<goals>
+							<goal>jar-no-fork</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+
+			<!-- -javadoc.jar (Maven Central requirement). Classpath mode (no module-info in
+			     this module), so it is clear of the JPMS module-mode javadoc trap. -->
+			<plugin>
+				<groupId>org.apache.maven.plugins</groupId>
+				<artifactId>maven-javadoc-plugin</artifactId>
+				<version>${javadoc.plugin.version}</version>
+				<configuration>
+					<source>17</source>
+					<failOnError>true</failOnError>
+					<doclint>all</doclint>
+				</configuration>
+				<executions>
+					<execution>
+						<id>attach-javadocs</id>
+						<goals>
+							<goal>jar</goal>
+						</goals>
+					</execution>
+				</executions>
+			</plugin>
+		</plugins>
+	</build>
+
+	<profiles>
+		<!-- Mirrors the core pom.xml `release` profile: GPG-sign the artifacts and route the
+		     deploy through the Central Publishing plugin. Activated by the publish-snapshot /
+		     publish-release CI jobs (`-P release ... deploy`). -->
+		<profile>
+			<id>release</id>
+			<build>
+				<plugins>
+					<plugin>
+						<groupId>org.apache.maven.plugins</groupId>
+						<artifactId>maven-gpg-plugin</artifactId>
+						<version>${gpg.plugin.version}</version>
+						<executions>
+							<execution>
+								<id>sign-artifacts</id>
+								<phase>verify</phase>
+								<goals><goal>sign</goal></goals>
+								<configuration>
+									<keyname>${gpg.keyname}</keyname>
+									<gpgArguments>
+										<arg>--pinentry-mode</arg>
+										<arg>loopback</arg>
+									</gpgArguments>
+								</configuration>
+							</execution>
+						</executions>
+					</plugin>
+
+					<plugin>
+						<groupId>org.sonatype.central</groupId>
+						<artifactId>central-publishing-maven-plugin</artifactId>
+						<version>${central.plugin.version}</version>
+						<extensions>true</extensions>
+						<configuration>
+							<publishingServerId>central</publishingServerId>
+							<autoPublish>true</autoPublish>
+							<waitUntil>published</waitUntil>
+						</configuration>
+					</plugin>
+				</plugins>
+			</build>
+		</profile>
+	</profiles>
+</project>
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaChatModel.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaChatModel.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaChatModel.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaChatModel.java
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModel.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModel.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModel.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModel.java
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaScoringModel.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaScoringModel.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaScoringModel.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaScoringModel.java
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaStreamingChatModel.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaStreamingChatModel.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/JllamaStreamingChatModel.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/JllamaStreamingChatModel.java
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/LangChain4jMapping.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/LangChain4jMapping.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/LangChain4jMapping.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/LangChain4jMapping.java
diff --git a/langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/package-info.java b/llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/package-info.java
similarity index 100%
rename from langchain4j-jllama/src/main/java/net/ladenthin/llama/langchain4j/package-info.java
rename to llama-langchain4j/src/main/java/net/ladenthin/llama/langchain4j/package-info.java
diff --git a/langchain4j-jllama/src/test/java/net/ladenthin/llama/langchain4j/JllamaChatModelIntegrationTest.java b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaChatModelIntegrationTest.java
similarity index 100%
rename from langchain4j-jllama/src/test/java/net/ladenthin/llama/langchain4j/JllamaChatModelIntegrationTest.java
rename to llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaChatModelIntegrationTest.java
diff --git a/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModelIntegrationTest.java b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModelIntegrationTest.java
new file mode 100644
index 00000000..79e0384a
--- /dev/null
+++ b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaEmbeddingModelIntegrationTest.java
@@ -0,0 +1,58 @@
+// SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
+//
+// SPDX-License-Identifier: MIT
+
+package net.ladenthin.llama.langchain4j;
+
+import static org.hamcrest.MatcherAssert.assertThat;
+import static org.hamcrest.Matchers.greaterThan;
+import static org.hamcrest.Matchers.is;
+import static org.hamcrest.Matchers.notNullValue;
+
+import dev.langchain4j.data.embedding.Embedding;
+import dev.langchain4j.data.segment.TextSegment;
+import dev.langchain4j.model.output.Response;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.Paths;
+import java.util.Arrays;
+import java.util.List;
+import net.ladenthin.llama.LlamaModel;
+import net.ladenthin.llama.parameters.ModelParameters;
+import org.junit.jupiter.api.Assumptions;
+import org.junit.jupiter.api.Test;
+
+/**
+ * End-to-end smoke test for {@link JllamaEmbeddingModel} over a real embedding model. Self-skips
+ * unless {@code -Dnet.ladenthin.llama.langchain4j.embedding.model=/abs/path/to/embedding.gguf}
+ * points at a GGUF loadable in embedding mode (and the native library is present), mirroring the
+ * core project's model-gated tests. CI reuses the already-cached nomic embedding model, so no extra
+ * download is introduced.
+ */
+class JllamaEmbeddingModelIntegrationTest {
+
+    private static Path modelPath() {
+        String path = System.getProperty("net.ladenthin.llama.langchain4j.embedding.model");
+        Assumptions.assumeTrue(path != null && !path.isEmpty(), "embedding model path property not set");
+        Path resolved = Paths.get(path);
+        Assumptions.assumeTrue(Files.exists(resolved), "embedding model file not present: " + resolved);
+        return resolved;
+    }
+
+    @Test
+    void embedsAllSegmentsInInputOrder() {
+        Path model = modelPath();
+        try (LlamaModel llama =
+                new LlamaModel(new ModelParameters().setModel(model.toString()).enableEmbedding())) {
+            JllamaEmbeddingModel embeddingModel = new JllamaEmbeddingModel(llama);
+
+            List<TextSegment> segments =
+                    Arrays.asList(TextSegment.from("hello world"), TextSegment.from("goodbye world"));
+            Response<List<Embedding>> response = embeddingModel.embedAll(segments);
+
+            assertThat(response, is(notNullValue()));
+            assertThat(response.content().size(), is(2));
+            assertThat(response.content().get(0).vector().length, is(greaterThan(0)));
+        }
+    }
+}
diff --git a/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaScoringModelIntegrationTest.java b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaScoringModelIntegrationTest.java
new file mode 100644
index 00000000..4c61c16d
--- /dev/null
+++ b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/JllamaScoringModelIntegrationTest.java
@@ -0,0 +1,57 @@
+// SPDX-FileCopyrightText: 2026 Bernard Ladenthin <bernard.ladenthin@gmail.com>
+//
+// SPDX-License-Identifier: MIT
+
+package net.ladenthin.llama.langchain4j;
+
+import static org.hamcrest.MatcherAssert.assertThat;
+import static org.hamcrest.Matchers.is;
+import static org.hamcrest.Matchers.notNullValue;
+
+import dev.langchain4j.data.segment.TextSegment;
+import dev.langchain4j.model.output.Response;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.Paths;
+import java.util.Arrays;
+import java.util.List;
+import net.ladenthin.llama.LlamaModel;
+import net.ladenthin.llama.parameters.ModelParameters;
+import org.junit.jupiter.api.Assumptions;
+import org.junit.jupiter.api.Test;
+
+/**
+ * End-to-end smoke test for {@link JllamaScoringModel} (re-ranker) over a real reranking model.
+ * Self-skips unless {@code -Dnet.ladenthin.llama.langchain4j.rerank.model=/abs/path/to/reranker.gguf}
+ * points at a GGUF loadable in reranking mode (and the native library is present), mirroring the
+ * core project's model-gated tests. CI reuses the already-cached jina reranker model, so no extra
+ * download is introduced.
+ */
+class JllamaScoringModelIntegrationTest {
+
+    private static Path modelPath() {
+        String path = System.getProperty("net.ladenthin.llama.langchain4j.rerank.model");
+        Assumptions.assumeTrue(path != null && !path.isEmpty(), "rerank model path property not set");
+        Path resolved = Paths.get(path);
+        Assumptions.assumeTrue(Files.exists(resolved), "rerank model file not present: " + resolved);
+        return resolved;
+    }
+
+    @Test
+    void scoresEverySegmentInInputOrder() {
+        Path model = modelPath();
+        try (LlamaModel llama =
+                new LlamaModel(new ModelParameters().setModel(model.toString()).enableReranking())) {
+            JllamaScoringModel scoringModel = new JllamaScoringModel(llama);
+
+            List<TextSegment> segments =
+                    Arrays.asList(
+                            TextSegment.from("A cat sat on the mat."),
+                            TextSegment.from("The stock market fell today."));
+            Response<List<Double>> response = scoringModel.scoreAll(segments, "domestic pets");
+
+            assertThat(response, is(notNullValue()));
+            assertThat(response.content().size(), is(2));
+        }
+    }
+}
diff --git a/langchain4j-jllama/src/test/java/net/ladenthin/llama/langchain4j/LangChain4jMappingTest.java b/llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/LangChain4jMappingTest.java
similarity index 100%
rename from langchain4j-jllama/src/test/java/net/ladenthin/llama/langchain4j/LangChain4jMappingTest.java
rename to llama-langchain4j/src/test/java/net/ladenthin/llama/langchain4j/LangChain4jMappingTest.java