diff --git a/sci-misc/llama-cpp/Manifest b/sci-misc/llama-cpp/Manifest
index 9f2d63aa5f..216381b579 100644
--- a/sci-misc/llama-cpp/Manifest
+++ b/sci-misc/llama-cpp/Manifest
@@ -6,3 +6,4 @@ DIST llama-cpp-0_pre7924.tar.gz 28899921 BLAKE2B b89c8c170d1c2d52390dfef35a4e085
 DIST llama-cpp-0_pre8198.tar.gz 29091642 BLAKE2B b5d02310a7e71fec88e29094aa7936cb520530fb9f81cef58f87c97948e6116734f9a19da5ad8daba1027f8e26308e5fa1f3e91e33a49539929f6e3475b08429 SHA512 bc8aef7cb29a32e0acf90d50cbf426c1a42d99659747203449ebf85f0c9c2fed98fd23b7e50b963dd537d3d6edecf17936d0cf7ae04266df4723e2d4a5c7ecfa
 DIST llama-cpp-0_pre8628.tar.gz 29611654 BLAKE2B 5efeb855b5833ca7a0ba552a142e9fac08773e0f0f31d9d3188c03840d91c91aa22edaf310a6472c9077cd4fb2a48c82371887374fb6d417dd022a18c51ff9e2 SHA512 7b86c37b3f926a0983fe6648629019e9f32001dbf474de9278d94d5e1a9e4c37dfaa9832557555724941c32c437efcbf2e91b0b0ad558936657e14ec5e52f8f0
 DIST llama-cpp-0_pre8838.tar.gz 33811834 BLAKE2B 1223a15237186dc8e11c454d76adfd875a9d6b3ee56c046fc0cf63415c7ebbb4b0dc2da49b7e697ecd9908dad6adcb4f2a5bd4665b7118d2c2508e3b215c1f39 SHA512 d345a2645dcad2564f94efdef45a7681b85f4079d30e68297b42a529ccfe4f50429067dce8685791c914cd59d00843c8e5d02c6ef7ce2d97bb990d6c896f5e9f
+DIST llama-cpp-0_pre9888.tar.gz 35116234 BLAKE2B 2ecc7b8a594441c83501ac91adf29182c1c308e74d52c57a0992b24526fab5639cf5844f3c04affa6864f576444b5c4aa2ed6f562927030d9144738dd56a3ad7 SHA512 c19994dad5415fe12d7215d16a4a32104b6320f0f004af86463b4c8f47e7529b3b32eac948ce6b3a3354c388a9d33db34d9d7d3d7f166f89130897567d441ce6
diff --git a/sci-misc/llama-cpp/llama-cpp-0_pre9442.ebuild b/sci-misc/llama-cpp/llama-cpp-0_pre9442.ebuild
new file mode 100644
index 0000000000..1dbc254a58
--- /dev/null
+++ b/sci-misc/llama-cpp/llama-cpp-0_pre9442.ebuild
@@ -0,0 +1,177 @@
+# Copyright 2026 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+ROCM_VERSION="6.3"
+
+inherit cmake cuda rocm linux-info
+
+TINY_LLAMAS_COMMIT="99dd1a73db5a37100bd4ae633f4cfce6560e1567"
+
+DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
+HOMEPAGE="https://github.com/ggml-org/llama.cpp"
+
+if [[ ${PV} == *9999* ]]; then
+	inherit git-r3
+	EGIT_REPO_URI="https://github.com/ggml-org/llama.cpp.git"
+else
+	MY_PV="b${PV#0_pre}"
+	SRC_URI="
+		https://github.com/ggml-org/llama.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz
+		webui? https://huggingface.co/buckets/ggml-org/llama-ui/resolve/b0
+	"
+	S="${WORKDIR}/llama.cpp-${MY_PV}"
+	KEYWORDS="~amd64"
+fi
+
+SRC_URI+="
+	examples? (
+		https://huggingface.co/ggml-org/tiny-llamas/resolve/${TINY_LLAMAS_COMMIT}/stories15M-q4_0.gguf
+			-> ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf
+	)
+"
+
+LICENSE="MIT"
+SLOT="0"
+CPU_FLAGS_X86=( avx avx2 f16c )
+
+# wwma USE explained here: https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md#hip
+IUSE="curl openblas +openmp blis rocm cuda opencl openssl vulkan flexiblas wmma examples"
+
+REQUIRED_USE="
+	?? (
+		openblas
+		blis
+		flexiblas
+	)
+	wmma? (
+		rocm
+	)
+"
+
+# curl is needed for pulling models from huggingface
+# numpy is used by convert_hf_to_gguf.py
+CDEPEND="
+	curl? ( net-misc/curl:= )
+	openblas? ( sci-libs/openblas:= )
+	openmp? ( llvm-runtimes/openmp:= )
+	blis? ( sci-libs/blis:= )
+	flexiblas? ( sci-libs/flexiblas:= )
+	rocm? (
+		>=dev-util/hip-${ROCM_VERSION}:=
+		>=sci-libs/hipBLAS-${ROCM_VERSION}:=
+		wmma? (
+			>=sci-libs/rocWMMA-${ROCM_VERSION}:=
+		)
+	)
+	cuda? ( dev-util/nvidia-cuda-toolkit:= )
+	openssl? ( dev-libs/openssl:= )
+"
+DEPEND="${CDEPEND}
+	opencl? ( dev-util/opencl-headers )
+	vulkan? (
+		dev-util/spirv-headers
+		dev-util/vulkan-headers
+	)
+"
+RDEPEND="${CDEPEND}
+	dev-python/numpy
+	opencl? ( dev-libs/opencl-icd-loader )
+	vulkan? ( media-libs/vulkan-loader )
+"
+BDEPEND="media-libs/shaderc"
+
+pkg_setup() {
+	if use rocm; then
+		linux-info_pkg_setup
+		if linux-info_get_any_version && linux_config_exists; then
+			if ! linux_chkconfig_present HSA_AMD_SVM; then
+				ewarn "To use ROCm/HIP, you need to have HSA_AMD_SVM option enabled in your kernel."
+			fi
+		fi
+	fi
+}
+
+src_prepare() {
+	use cuda && cuda_src_prepare
+	cmake_src_prepare
+	if use examples; then
+		mkdir -p "${BUILD_DIR}/tinyllamas" || die
+		cp "${DISTDIR}/ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf" \
+			"${BUILD_DIR}/tinyllamas/stories15M-q4_0.gguf" || die
+	fi
+}
+
+src_configure() {
+	local mycmakeargs=(
+		-DLLAMA_BUILD_TESTS=OFF
+		-DLLAMA_BUILD_EXAMPLES=$(usex examples)
+		-DLLAMA_BUILD_SERVER=ON
+		-DCMAKE_SKIP_BUILD_RPATH=ON
+		-DGGML_NATIVE=0	# don't set march
+		-DGGML_RPC=ON
+		-DLLAMA_CURL=$(usex curl)
+		-DLLAMA_OPENSSL=$(usex openssl)
+		-DBUILD_NUMBER="1"
+		-DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON
+		-DGGML_CUDA=$(usex cuda)
+		-DGGML_OPENCL=$(usex opencl)
+		-DGGML_OPENMP=$(usex openmp)
+		-DGGML_VULKAN=$(usex vulkan)
+
+		# avoid clashing with whisper.cpp
+		-DCMAKE_INSTALL_LIBDIR="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
+		-DCMAKE_INSTALL_RPATH="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
+	)
+
+	if use webui ; then
+
+		mycmakeargs+=(
+			-DLLAMA_USE_PREBUILT_UI=ON
+		)
+	fi
+
+	if use openblas ; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+		)
+	fi
+
+	if use blis ; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME
+		)
+	fi
+
+	if use flexiblas; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FlexiBLAS
+		)
+	fi
+
+	if use cuda; then
+		local -x CUDAHOSTCXX="$(cuda_gccdir)"
+		# tries to recreate dev symlinks
+		cuda_add_sandbox
+		addpredict "/dev/char/"
+	fi
+
+	if use rocm; then
+		rocm_use_hipcc
+		mycmakeargs+=(
+			-DGGML_HIP=ON -DAMDGPU_TARGETS=$(get_amdgpu_flags)
+			-DGGML_HIP_ROCWMMA_FATTN=$(usex wmma)
+		)
+	fi
+
+	cmake_src_configure
+}
+
+src_install() {
+	cmake_src_install
+	dobin "${BUILD_DIR}/bin/rpc-server"
+
+	# avoid clashing with whisper.cpp
+	rm -rf "${ED}/usr/include"
+}