diff --git a/sci-misc/llama-cpp/Manifest b/sci-misc/llama-cpp/Manifest
new file mode 100644
index 0000000..e03cdce
--- /dev/null
+++ b/sci-misc/llama-cpp/Manifest
@@ -0,0 +1,3 @@
+DIST ggml-org_models_tinyllamas_stories15M-q4_0-99dd1a73db5a37100bd4ae633f4cfce6560e1567.gguf 19077344 SHA256 6151b1929d7f5aa3385d9ddef3393e55587c0a55de661562322bc51dfda93a04 SHA512 f9944886089958e0d97b1906cfd45020e0821c65429346e76fae29136c634ae5d039dffbae5933a95b0674f4acd87b656feb9f9e1b16dd434c5c9b5886f4f617 WHIRLPOOL 16d7f91e83fd02c9e9a9babb8d21e4162834d11b462da2131b816386f8a9bf0012ae7b4e6a315734cb2683208d29f6f1f3da7c1deca5c405cfa28347e3f0ff1d
+DIST llama-cpp-0_pre9763.tar.gz 34971715 SHA256 31c10b6ca19f6e02fc5557721da3085b512c0a9af1707f7843a9f882fddbe49c SHA512 d33dcfca09603a426db084e89d5e8bf65d2af8eec3f73b7300cdd4a93df1716c1974df3a4abddc9a627202a245e9be982a3111a0725bb5898b41bf9270c240ac WHIRLPOOL c9e69e29fd3563b1e691f082b611369aa4781a6d91c589e13584c144006d2d8523773e74dcf1b0c5a28e14f07316dd0a4346b42194e780cc63bc65b802a77bd6
+EBUILD llama-cpp-0_pre9763.ebuild 3827 SHA256 f83eb725a5d1f8c8313b1a6fdf72ad893bd1bf93f7ea64c15c620d71f4ef68f3 SHA512 0ddee1d9c023966730592c563d9ee70ce7cf080aa65754d28efc518bc5d1ac333b3599351858ecb9e9a18d4ab8a0c84d8c02f80dc21c291c08c4ac1e022c2384 WHIRLPOOL 6e243f01ed039b7fc3c5f8f53101d364c9d2fbd6f29b58b003348c2e53725ad70c8b38326620c606723f287751461c3ad0e71debee1fc9cf2ff41a15eb71419f
diff --git a/sci-misc/llama-cpp/llama-cpp-0_pre9763.ebuild b/sci-misc/llama-cpp/llama-cpp-0_pre9763.ebuild
new file mode 100644
index 0000000..2295b3d
--- /dev/null
+++ b/sci-misc/llama-cpp/llama-cpp-0_pre9763.ebuild
@@ -0,0 +1,167 @@
+# 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 -> ${P}.tar.gz"
+	S="${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/$(get_libdir)/llama.cpp"
+	)
+
+	if use openblas ; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+		)
+	fi
+
+	if use blis ; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FLAME
+		)
+	fi
+
+	if use flexiblas; then
+		mycmakeargs+=(
+			-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=FlexiBLAS
+		)
+	fi
+
+	if use cuda; then
+		local -x CUDAHOSTCXX="$(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"
+}