diff --git a/pkgs/by-name/ol/ollama/package.nix b/pkgs/by-name/ol/ollama/package.nix
index cc0520fad3f1..4181f137fe8d 100644
--- a/pkgs/by-name/ol/ollama/package.nix
+++ b/pkgs/by-name/ol/ollama/package.nix
@@ -7,7 +7,6 @@
   stdenv,
   addDriverRunpath,
   nix-update-script,
-  coreutils,
 
   cmake,
   gitMinimal,
@@ -108,6 +107,17 @@ let
 
   cudaPath = lib.removeSuffix "-${cudaMajorVersion}" cudaToolkit;
 
+  # Since v0.30, llama.cpp is consumed via CMake FetchContent rather than
+  # vendored in-tree. Pre-stage the pinned commit (read from upstream's
+  # `LLAMA_CPP_VERSION` file — currently `b9493`) so the FetchContent step
+  # uses our copy instead of trying to clone over the network in the sandbox.
+  llamaCppSrc = fetchFromGitHub {
+    owner = "ggml-org";
+    repo = "llama.cpp";
+    rev = "a731805cedc83c0514cbd808a2e38ec46c759cc2"; # tag b9493
+    hash = "sha256-DO9J1mx9Jlp6qtCiJp2ZEi6R7H2YX1/sD7DGgBCtt0U=";
+  };
+
   wrapperOptions = [
     # ollama embeds llama-cpp binaries which actually run the ai models
     # these llama-cpp binaries are unaffected by the ollama binary's DT_RUNPATH
@@ -132,25 +142,25 @@ let
     if enableCuda then
       buildGoModule.override { stdenv = cudaPackages.backendStdenv; }
     else if enableRocm then
-      buildGoModule.override { stdenv = rocmPackages.stdenv; }
+      buildGoModule.override { inherit (rocmPackages) stdenv; }
     else if enableVulkan then
-      buildGoModule.override { stdenv = vulkan-tools.stdenv; }
+      buildGoModule.override { inherit (vulkan-tools) stdenv; }
     else
       buildGoModule;
   inherit (lib) licenses platforms maintainers;
 in
 goBuild (finalAttrs: {
   pname = "ollama";
-  version = "0.24.0";
+  version = "0.30.4";
 
   src = fetchFromGitHub {
     owner = "ollama";
     repo = "ollama";
     tag = "v${finalAttrs.version}";
-    hash = "sha256-cSZtbF0oUI7a++baTipF6OUu+w8tBpILzE0Wm1Y6wUk=";
+    hash = "sha256-IHX8o1Ty4Sdht5YeUYLnNPjOV7O95WeNng/coO+MHS8=";
   };
 
-  vendorHash = "sha256-Lc1Ktdqtv2VhJQssk8K1UOimeEjVNvDWePE9WkamCos=";
+  vendorHash = "sha256-lZdGzGb9xRjTm1Rm7/wHjqM490gLznLEndmb4mNbCX0=";
   proxyVendor = true;
 
   env =
@@ -190,25 +200,31 @@ goBuild (finalAttrs: {
     ++ lib.optionals enableVulkan vulkanLibs;
 
   # replace inaccurate version number with actual release version
-  # and replace core utils tools from their FHS location to nix store
   postPatch = ''
     substituteInPlace version/version.go \
       --replace-fail 0.0.0 '${finalAttrs.version}'
-    substituteInPlace cmd/launch/openclaw_test.go \
-      --replace-fail '/bin/mkdir' '${coreutils}/bin/mkdir' \
-      --replace-fail '/bin/cat' '${coreutils}/bin/cat' \
-      --replace-fail '/usr/bin/env' '${coreutils}/bin/env' \
-      --replace-fail '/usr/bin/sort' '${coreutils}/bin/sort' \
-      --replace-fail '/bin/chmod' '${coreutils}/bin/chmod'
-    substituteInPlace cmd/launch/hermes_test.go \
-      --replace-fail '/bin/mkdir' '${coreutils}/bin/mkdir' \
-      --replace-fail '/bin/cat' '${coreutils}/bin/cat' \
-      --replace-fail '/bin/chmod' '${coreutils}/bin/chmod'
-    substituteInPlace cmd/launch/kimi_test.go \
-       --replace-fail '/bin/mkdir' '${coreutils}/bin/mkdir' \
-       --replace-fail '/bin/cat' '${coreutils}/bin/cat' \
-       --replace-fail '/bin/chmod' '${coreutils}/bin/chmod'
+
+    # cmd/launch/*_test.go are integration tests for user-facing CLI
+    # launchers (claude, qwen, cline, codex, kimi, droid, openclaw, hermes,
+    # …) that install the target binary via npm and then exec it on PATH.
+    # Both prerequisites are unavailable in the nix sandbox, so the launch
+    # subpackage's tests can't pass here. Drop them.
+    rm cmd/launch/*_test.go
+
     rm -r app
+
+    # Pre-stage llama.cpp for the FetchContent step and apply Ollama's
+    # compat patch. When FETCHCONTENT_SOURCE_DIR_LLAMA_CPP is set, neither
+    # `cmake/local.cmake` nor `llama/server/CMakeLists.txt` auto-applies
+    # the patch (the parent's ExternalProject_Add passes
+    # OLLAMA_LLAMA_CPP_SKIP_COMPAT_PATCH=ON to the child build) — the
+    # caller has to. The apply-patch.cmake script is idempotent so this
+    # is safe to re-run.
+    cp -r ${llamaCppSrc} $TMPDIR/llama-cpp-src
+    chmod -R +w $TMPDIR/llama-cpp-src
+    ( cd $TMPDIR/llama-cpp-src && \
+      cmake -DPATCH_DIR=$NIX_BUILD_TOP/source/llama/compat \
+        -P $NIX_BUILD_TOP/source/llama/compat/apply-patch.cmake )
   ''
   # disable tests that fail in sandbox due to Metal init failure
   + lib.optionalString stdenv.hostPlatform.isDarwin ''
@@ -217,12 +233,10 @@ goBuild (finalAttrs: {
     rm model/models/nemotronh/model_omni_test.go
   '';
 
-  overrideModAttrs = (
-    finalAttrs: prevAttrs: {
-      # don't run llama.cpp build in the module fetch phase
-      preBuild = "";
-    }
-  );
+  overrideModAttrs = _: _: {
+    # don't run llama.cpp build in the module fetch phase
+    preBuild = "";
+  };
 
   preBuild =
     let
@@ -236,20 +250,53 @@ goBuild (finalAttrs: {
       cudaArchitectures = builtins.concatStringsSep ";" (map removeSMPrefix cudaArches);
       rocmTargets = builtins.concatStringsSep ";" rocmGpuTargets;
 
+      # Since 0.30, Ollama splits the llama.cpp build into per-accelerator
+      # "runners" gated by OLLAMA_LLAMA_BACKENDS. Without setting it the
+      # build silently produces only the CPU runner — ollama-cuda would
+      # ship without `libggml-cuda.so` and fall back to CPU at runtime.
+      # The accepted values map to cmake/local.cmake's elseif chain
+      # (cuda_v12 / cuda_v13 / rocm_v7_1 / rocm_v7_2 / vulkan / cuda_jetpack*).
+      rocmMajorVersion = lib.versions.major rocmPackages.clr.version;
+      rocmMinorVersion = lib.versions.minor rocmPackages.clr.version;
+      llamaBackend =
+        if enableCuda then
+          "cuda_v${cudaMajorVersion}"
+        else if enableRocm then
+          "rocm_v${rocmMajorVersion}_${rocmMinorVersion}"
+        else if enableVulkan then
+          "vulkan"
+        else
+          "";
+
       cmakeFlagsCudaArchitectures = lib.optionalString enableCuda "-DCMAKE_CUDA_ARCHITECTURES='${cudaArchitectures}'";
       cmakeFlagsRocmTargets = lib.optionalString enableRocm "-DAMDGPU_TARGETS='${rocmTargets}'";
+      cmakeFlagsBackend = lib.optionalString (
+        llamaBackend != ""
+      ) "-DOLLAMA_LLAMA_BACKENDS=${llamaBackend}";
 
     in
     ''
       cmake -B build \
         -DCMAKE_SKIP_BUILD_RPATH=ON \
         -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+        -DFETCHCONTENT_SOURCE_DIR_LLAMA_CPP="$TMPDIR/llama-cpp-src" \
         ${cmakeFlagsCudaArchitectures} \
         ${cmakeFlagsRocmTargets} \
+        ${cmakeFlagsBackend}
 
       cmake --build build -j $NIX_BUILD_CORES
     '';
 
+  # The llama.cpp sub-build is driven by ExternalProject_Add and does
+  # not inherit the parent's CMAKE_SKIP_BUILD_RPATH setting, so its
+  # `.so` payloads end up with build-dir entries in RPATH. Drop them
+  # before the forbidden-references check. $ORIGIN is preserved
+  # unconditionally; only absolute /nix/store entries are kept.
+  preFixup = ''
+    find $out/lib/ollama -type f \( -name '*.so' -o -name '*.so.*' \) \
+      -exec patchelf --shrink-rpath --allowed-rpath-prefixes /nix/store {} +
+  '';
+
   # ollama looks for acceleration libs in ../lib/ollama/ (now also for CPU-only with arch specific optimizations)
   # https://github.com/ollama/ollama/blob/v0.21.1/docs/development.md#library-detection
   postInstall = ''