diff --git a/sci-misc/llama-cpp/Manifest b/sci-misc/llama-cpp/Manifest
index 715b79116a..e32bfc7779 100644
--- a/sci-misc/llama-cpp/Manifest
+++ b/sci-misc/llama-cpp/Manifest
@@ -2,6 +2,8 @@ DIST ggml-org_models_tinyllamas_stories15M-q4_0-99dd1a73db5a37100bd4ae633f4cfce6
 DIST llama-b10621-ui.tar.gz 3082196 BLAKE2B 92a794b6c971737cb85d5aafadf59a07faf7d78d77d8fd8eb86412add84317be0f38d1fab540f4c1786269b399b575c61518e80bb4b98b6e2792ead3d49ae833 SHA512 76d1c53d84a2588055ce2066fee5a8f2ea8df37384a1d16456d90a09f38a290df860b3a18910b1b512ea6fdba5842ce0b2513bbc3cb6de70cb856584fabfe209
 DIST llama-b10780-ui.tar.gz 3084093 BLAKE2B 92caf7e64a1f40367687ec6e48a2eb2ac90e44983ec7a169397e1b7d085bafa52d37788a2f98707ac10ffd2b5b85e28cad3ab929b119a73f05cb1a6c64eb1dfd SHA512 89d15d3bbfad6b26d68896fc19ec6e05d947e61be8b1c836a2e187ff9dc54595abc8b74532d002cbf6d6ee1c89ac4c08e2261233e51924a35621662278755341
 DIST llama-b10809-ui.tar.gz 3084118 BLAKE2B 3cd4e0bd708fb673507b1480e7e814bcb25b84ca096c22d491fbb02ea28523f28517ad0cbf15a7c689801599206963b1268239eba6d028a9343956710c39d07c SHA512 485b4331f547bcf53fa40691e43ca97de90301f656b8176c2ae5091bb2c6d0dcee57439f12a02a06fa5a397f5aba2481e8614c2a06da1396fa8efa58b704c973
+DIST llama-b10905-ui.tar.gz 3083354 BLAKE2B 6fe9ee8e53e6d2c70f57e076fa3b75ea1e81c351612d95ec5e4bcb8e87fa8d77516cf635eac783eeefd6013d64137f05dc8676e5539539dcb8c6cbd9cc087dfa SHA512 f6a49e3da8cca1036e4669ce9bdd69bfd8e112bbf462ac0ea0ac28266834ac3525ec779c0ff2c16a3e4e2f1d9d9dd49ce1cd1a9132cf5f5f80a22ba0d0362b86
 DIST llama-cpp-0.3.0.tar.gz 36951021 BLAKE2B fa59feff83100b137710aeed91ffe721d057f891749aff8c0f5758c86dd22760d0456f808137aad7ee2faf8e14956f603dd6d0bbb44a0767f026c61a94ad2d30 SHA512 c92095a186920110cf9f542d00b004236d92f1ab3f22267bb6ed8fc4b65dc1c32b2705efab202849756d464ed2213c6ee4d5613482f56dad26ae6eb3ba16a91d
 DIST llama-cpp-0.3.0_p10780.tar.gz 37148504 BLAKE2B c65bbbe9c15e79cb80fec728246e0e16b0fe52bd8523e3f948f5eea7e02b0d8762eb02dfc86cc44bf38180398ca8860786b637b7139e3afd9367160467bbade5 SHA512 26392bc161594dccfd74287e5664cc56c14222f7281e10129e7980f48275368ff003dffb36b8b7d742c4aabcec2562c9aaa78d1abc533f618fc5aab8d66c8575
 DIST llama-cpp-0.4.0.tar.gz 37288249 BLAKE2B a0cd0c8251b4dc32681426e323e52b03c9a733a3d29d8e733334702c2082b2aff664906811dee0298ea950a28fec40771e5c27aa5a2c32b7f7b9b9cab512249d SHA512 12592516c0ac70ed53d64fc0c9f82bbb156defa507c7364f1a753eb29f3498e0e13a282e139c2eff7d2eb00921b4ef107b2fa90a5b66f8a3c45409d4fcb0834f
+DIST llama-cpp-0.4.0_p10905.tar.gz 37373261 BLAKE2B e966b0a9abee8486b197a35958c2a4adaaffc4911595c9372da46d59e75ee70f1609e63bd79078ea32b61d54d763eeab6a05e65f7d8cd0cf6916708fe97cb717 SHA512 c2f33fdc62c9e43e643d35128688b75c9227416cb99fc4796b992f759e340ddd16ad7a8ca43b31d3214536e2b9bf1f24f0cdbc4eef39793f2b951c99775a40e5
diff --git a/sci-misc/llama-cpp/llama-cpp-0.4.0_p10905.ebuild b/sci-misc/llama-cpp/llama-cpp-0.4.0_p10905.ebuild
new file mode 100644
index 0000000000..c1a5838e44
--- /dev/null
+++ b/sci-misc/llama-cpp/llama-cpp-0.4.0_p10905.ebuild
@@ -0,0 +1,238 @@
+# Copyright 2026 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+ROCM_VERSION="7.1"
+
+inherit cmake cuda rocm linux-info toolchain-funcs
+
+TINY_LLAMAS_COMMIT="99dd1a73db5a37100bd4ae633f4cfce6560e1567"
+
+DESCRIPTION="Port of Facebook's LLaMA model in C/C++"
+HOMEPAGE="https://github.com/ggml-org/llama.cpp"
+
+if [[ ${PV} == *9999* ]]; then
+	inherit git-r3
+	EGIT_REPO_URI="https://github.com/ggml-org/llama.cpp.git"
+	LLAMA_BUILD_IS_DEV=ON
+	LLAMA_BUILD_NUMBER="b0000"
+	LLAMA_UI_VERSION="b10905"
+	MY_PV="${PV}"
+else
+	if [[ $(ver_cut 4) == "p" ]]; then
+		LLAMA_BUILD_IS_DEV=ON
+		LLAMA_BUILD_NUMBER="b$(ver_cut 5)"
+		MY_PV="${LLAMA_BUILD_NUMBER}"
+		S="${WORKDIR}/llama.cpp-${MY_PV}"
+	else
+		LLAMA_BUILD_IS_DEV=OFF
+		# Set from https://github.com/ggml-org/llama.cpp/releases/download/v${PV}/nightly-tag.txt
+		LLAMA_BUILD_NUMBER="b10905"
+		MY_PV="v${PV}"
+		S="${WORKDIR}/llama.cpp-${PV}"
+	fi
+	LLAMA_UI_VERSION="${LLAMA_BUILD_NUMBER}"
+	SRC_URI="https://github.com/ggml-org/llama.cpp/archive/refs/tags/${MY_PV}.tar.gz -> ${P}.tar.gz"
+	KEYWORDS="~amd64"
+fi
+
+SRC_URI+="
+	https://github.com/ggml-org/llama.cpp/releases/download/${LLAMA_UI_VERSION}/llama-${LLAMA_UI_VERSION}-ui.tar.gz
+	examples? (
+		https://huggingface.co/ggml-org/tiny-llamas/resolve/${TINY_LLAMAS_COMMIT}/stories15M-q4_0.gguf
+			-> ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf
+	)
+"
+
+LICENSE="MIT"
+SLOT="0"
+X86_CPU_FLAGS=(
+	amx_bf16
+	amx_int8
+	amx_tile
+	avx
+	avx2
+	avx512_bf16
+	avx512_vnni
+	avx512f
+	avx512vbmi
+	avx_vnni
+	bmi2
+	f16c
+	fma3
+	sse4_2
+)
+CPU_FLAGS=( "${X86_CPU_FLAGS[@]/#/cpu_flags_x86_}" )
+
+# wwma USE explained here: https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md#hip
+IUSE="${CPU_FLAGS[*]} curl openblas +openmp blis rocm cuda opencl openssl vulkan flexiblas wmma examples"
+
+REQUIRED_USE="
+	?? (
+		openblas
+		blis
+		flexiblas
+	)
+	rocm? ( ${ROCM_REQUIRED_USE} )
+	wmma? (
+		rocm
+	)
+"
+
+# curl is needed for pulling models from huggingface
+# numpy is used by convert_hf_to_gguf.py
+CDEPEND="
+	curl? ( net-misc/curl:= )
+	openblas? ( sci-libs/openblas:= )
+	openmp? ( llvm-runtimes/openmp:= )
+	blis? ( sci-libs/blis:= )
+	flexiblas? ( sci-libs/flexiblas:= )
+	rocm? (
+		>=dev-util/hip-${ROCM_VERSION}:=
+		>=sci-libs/hipBLAS-${ROCM_VERSION}:=[${ROCM_USEDEP}]
+		wmma? (
+			>=sci-libs/rocWMMA-${ROCM_VERSION}:=[${ROCM_USEDEP}]
+		)
+	)
+	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_pretend() {
+	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
+}
+
+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
+
+	[[ ${MERGE_TYPE} != binary ]] && use openmp && tc-check-openmp
+}
+
+src_unpack() {
+	[[ ${PV} == *9999* ]] && git-r3_src_unpack
+	default
+}
+
+src_prepare() {
+	use cuda && cuda_src_prepare
+	cmake_src_prepare
+	if use examples; then
+		mkdir -p "${BUILD_DIR}/tinyllamas" || die
+		cp "${DISTDIR}/ggml-org_models_tinyllamas_stories15M-q4_0-${TINY_LLAMAS_COMMIT}.gguf" \
+			"${BUILD_DIR}/tinyllamas/stories15M-q4_0.gguf" || die
+	fi
+
+	# Move web UI assets where they belong, bug #979245
+	mkdir -p "${S}"/tools/ui || die
+	cp -a "${WORKDIR}"/llama-${LLAMA_UI_VERSION} "${S}"/tools/ui/dist || die
+}
+
+src_configure() {
+	local mycmakeargs=(
+		-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
+		-DLLAMA_BUILD_IS_DEV=${LLAMA_BUILD_IS_DEV}
+		-DLLAMA_BUILD_TESTS=OFF
+		-DLLAMA_BUILD_EXAMPLES=$(usex examples)
+		-DLLAMA_BUILD_SERVER=ON
+		-DBUILD_NUMBER="${LLAMA_BUILD_NUMBER}"
+		-DLLAMA_CURL=$(usex curl)
+		-DLLAMA_OPENSSL=$(usex openssl)
+		-DGENTOO_REMOVE_CMAKE_BLAS_HACK=ON
+
+		# avoid clashing with whisper.cpp
+		-DCMAKE_INSTALL_LIBDIR="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
+		-DCMAKE_INSTALL_RPATH="${EPREFIX}/usr/$(get_libdir)/llama.cpp"
+	)
+
+	# GGML backends
+	mycmakeargs+=(
+		-DGGML_NATIVE=OFF	# don't set march
+
+		# CPU Flags
+		-DGGML_SSE42=$(usex cpu_flags_x86_sse4_2)
+		-DGGML_AVX=$(usex cpu_flags_x86_avx)
+		-DGGML_AVX_VNNI=$(usex cpu_flags_x86_avx_vnni)
+		-DGGML_AVX2=$(usex cpu_flags_x86_avx2)
+		-DGGML_BMI2=$(usex cpu_flags_x86_bmi2)
+		-DGGML_AVX512=$(usex cpu_flags_x86_avx512f)
+		-DGGML_AVX512_VBMI=$(usex cpu_flags_x86_avx512vbmi)
+		-DGGML_AVX512_VNNI=$(usex cpu_flags_x86_avx512_vnni)
+		-DGGML_AVX512_BF16=$(usex cpu_flags_x86_avx512_bf16)
+		-DGGML_FMA=$(usex cpu_flags_x86_fma3)
+		-DGGML_F16C=$(usex cpu_flags_x86_f16c)
+		-DGGML_AMX_TILE=$(usex cpu_flags_x86_amx_tile)
+		-DGGML_AMX_INT8=$(usex cpu_flags_x86_amx_int8)
+		-DGGML_AMX_BF16=$(usex cpu_flags_x86_amx_bf16)
+
+		-DGGML_CUDA=$(usex cuda)
+		-DGGML_VULKAN=$(usex vulkan)
+		-DGGML_OPENMP=$(usex openmp)
+		-DGGML_RPC=ON
+		-DGGML_OPENCL=$(usex opencl)
+	)
+
+	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/ggml-rpc-server"
+
+	# avoid clashing with whisper.cpp
+	rm -r "${ED}/usr/include" || die
+}