diff --git a/misc/ollama/Makefile b/misc/ollama/Makefile
index aba38aaf028e..7c41818fd3c1 100644
--- a/misc/ollama/Makefile
+++ b/misc/ollama/Makefile
@@ -1,274 +1,285 @@
 PORTNAME=	ollama
 DISTVERSIONPREFIX=	v
-DISTVERSION=	0.32.5
-PORTREVISION=	3
+DISTVERSION=	0.34.0
 CATEGORIES=	misc # machine-learning
 
 MAINTAINER=	yuri@FreeBSD.org
 COMMENT=	Run text/image LLMs and programming APIs locally and in the cloud
 WWW=		https://ollama.com \
 		https://github.com/ollama/ollama
 
 LICENSE=	MIT
 LICENSE_FILE=	${WRKSRC}/LICENSE
 
 ONLY_FOR_ARCHS=		amd64
 ONLY_FOR_ARCHS_REASON=	heavily relies on SIMD, no HW to test on arm and other architectures, also errors out: Unsupported ARM target OS: FreeBSD # patches are welcome
 
 BROKEN_i386=	fails to compile: x/mlxrunner/mlx/memory.go:40:11: 1 << (4 * 10) (untyped int constant 1099511627776) overflows int
 
 BUILD_DEPENDS=	bash:shells/bash \
 		${LOCALBASE}/include/miniaudio/miniaudio.h:audio/miniaudio \
 		${LOCALBASE}/include/nlohmann/json_fwd.hpp:devel/nlohmann-json \
 		${LOCALBASE}/include/stb/stb_image.h:devel/stb \
 		patchelf:sysutils/patchelf
 
 USES=		cmake:indirect go:1.26+,modules localbase pkgconfig
 USE_LDCONFIG=	${PREFIX}/lib/ollama ${PREFIX}/lib/ollama/vulkan
 USE_RC_SUBR=	ollama
 
+USE_GITHUB=	nodefault
+GH_TUPLE=	mlc-ai:xgrammar:v0.2.5:xgrammar/xgrammar
+
 GO_MODULE=	github.com/yurivict/${PORTNAME} # fork with FreeBSD patches
 GO_TARGET=	.
 GO_ENV+=	CGO_CXXFLAGS="${CXXFLAGS}"
 
-LLAMA_CPP_VERSION=	b10091	# from the LLAMA_CPP_VERSION file in the llama.cpp repo
-GGML_SO_VERSION=	0.17.0	# tied to LLAMA_CPP_VERSION; update when llama.cpp changes
+LLAMA_CPP_VERSION=	b10760	# from the LLAMA_CPP_VERSION file in the llama.cpp repo
+LLAMA_CPP_SO_VERSION=	0.3.0
+GGML_SO_VERSION=	0.22.0	# tied to LLAMA_CPP_VERSION; update when llama.cpp changes
 MLX_CORE_VERSION=	0.31.2
 MLX_C_VERSION=		fba4470b89073180056c9ea46c443051375f7399
 JSON_VERSION=		3.11.3
 
 DISTFILES+=	${LLAMA_CPP_VERSION}.tar.gz:llamasrc \
 		llama-${LLAMA_CPP_VERSION}-ui.tar.gz:llamasrc
 MASTER_SITES+=	https://github.com/ggml-org/llama.cpp/archive/refs/tags/:llamasrc \
 		https://github.com/ggml-org/llama.cpp/releases/download/${LLAMA_CPP_VERSION}/:llamasrc
 
 PLIST_FILES=	bin/${PORTNAME} \
 		bin/ollama-limit-gpu-layers
 
 OPTIONS_GROUP=		BACKENDS
 OPTIONS_GROUP_BACKENDS=	CPU VULKAN MLX
-OPTIONS_DEFAULT=	CPU MLX
+OPTIONS_DEFAULT=	CPU VULKAN MLX
 
 CPU_DESC=		Build CPU backend shared libraries for various SIMD instruction sets
 CPU_PLIST_FILES=	lib/ollama/llama-server \
 			lib/ollama/llama-quantize \
 			lib/ollama/libggml-base.so \
 			lib/ollama/libggml-base.so.0 \
 			lib/ollama/libggml-base.so.${GGML_SO_VERSION} \
 			lib/ollama/libggml.so \
 			lib/ollama/libggml.so.0 \
 			lib/ollama/libggml.so.${GGML_SO_VERSION} \
 			lib/ollama/libllama-common.so \
-			lib/ollama/libllama-common.so.0 \
-			lib/ollama/libllama-common.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+			lib/ollama/libllama-common.so.${LLAMA_CPP_SO_VERSION:R:R} \
+			lib/ollama/libllama-common.so.${LLAMA_CPP_SO_VERSION} \
 			lib/ollama/libllama-quantize-impl.so \
 			lib/ollama/libllama-server-impl.so \
 			lib/ollama/libllama.so \
-			lib/ollama/libllama.so.0 \
-			lib/ollama/libllama.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+			lib/ollama/libllama.so.${LLAMA_CPP_SO_VERSION:R:R} \
+			lib/ollama/libllama.so.${LLAMA_CPP_SO_VERSION} \
 			lib/ollama/libmtmd.so \
-			lib/ollama/libmtmd.so.0 \
-			lib/ollama/libmtmd.so.0.0.${LLAMA_CPP_VERSION:S/b//}
+			lib/ollama/libmtmd.so.${LLAMA_CPP_SO_VERSION:R:R} \
+			lib/ollama/libmtmd.so.${LLAMA_CPP_SO_VERSION}
 .if ${MACHINE_ARCH} == "amd64" || ${MACHINE_ARCH} == "i386"
 CPU_PLIST_FILES+=	lib/ollama/libggml-cpu-alderlake.so \
 			lib/ollama/libggml-cpu-cannonlake.so \
 			lib/ollama/libggml-cpu-cascadelake.so \
 			lib/ollama/libggml-cpu-cooperlake.so \
 			lib/ollama/libggml-cpu-haswell.so \
 			lib/ollama/libggml-cpu-icelake.so \
 			lib/ollama/libggml-cpu-ivybridge.so \
 			lib/ollama/libggml-cpu-piledriver.so \
 			lib/ollama/libggml-cpu-sandybridge.so \
 			lib/ollama/libggml-cpu-sapphirerapids.so \
 			lib/ollama/libggml-cpu-skylakex.so \
 			lib/ollama/libggml-cpu-sse42.so \
 			lib/ollama/libggml-cpu-x64.so \
 			lib/ollama/libggml-cpu-zen4.so
 .endif
 
 VULKAN_DESC=		Build Vulkan GPU backend shared library
 VULKAN_BUILD_DEPENDS=	glslc:graphics/shaderc \
 			${LOCALBASE}/include/vulkan/vulkan.h:graphics/vulkan-headers \
 			${LOCALBASE}/share/cmake/SPIRV-Headers/SPIRV-HeadersConfig.cmake:graphics/spirv-headers
 VULKAN_LIB_DEPENDS=	libvulkan.so:graphics/vulkan-loader
 VULKAN_PLIST_FILES=	lib/ollama/vulkan/libggml-vulkan.so
-VULKAN_BROKEN=		shaderc doesn't support latest features of glslc, see https://github.com/google/shaderc/issues/1590
 
 MLX_DESC=		Build MLX backend for image generation (CPU)
 MLX_BUILD_DEPENDS=	${LOCALBASE}/lib/cmake/fmt/fmt-config.cmake:devel/libfmt
 MLX_LIB_DEPENDS=	libopenblas.so:math/openblas
 MLX_PLIST_FILES=	lib/ollama/libmlx.so \
 			lib/ollama/libmlxc.so
 
 _CMAKE_FLAGS=	-DCMAKE_BUILD_TYPE=Release \
 		-DGGML_BACKEND_DL=ON \
 		-DGGML_BACKEND_DIR=${PREFIX}/lib/ollama \
 		-DFETCHCONTENT_FULLY_DISCONNECTED=ON \
 		-DFETCHCONTENT_SOURCE_DIR_LLAMA_CPP=${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION} \
 		-DOLLAMA_LLAMA_CPP_SKIP_COMPAT_PATCH=ON \
 		-DLLAMA_BUILD_NUMBER=${LLAMA_CPP_VERSION:S/b//}
 
 .include <bsd.port.options.mk>
 
 .if ${PORT_OPTIONS:MVULKAN}
 _CMAKE_FLAGS+=	-DOLLAMA_LLAMA_BACKENDS=vulkan
 .endif
 
 .if ${PORT_OPTIONS:MMLX}
 GO_BUILDFLAGS+=	-tags mlx
 DISTFILES+=	v${MLX_CORE_VERSION}.tar.gz:mlxsrc \
 		${MLX_C_VERSION}.tar.gz:mlxcsrc \
 		json.tar.xz:jsonsrc
 MASTER_SITES+=	https://github.com/ml-explore/mlx/archive/refs/tags/:mlxsrc \
 		https://github.com/ml-explore/mlx-c/archive/:mlxcsrc \
 		https://github.com/nlohmann/json/releases/download/v${JSON_VERSION}/:jsonsrc
 .endif
 
 post-extract:
 	# Copy pre-built UI assets into llama.cpp source tree so the build
 	# can embed them without running npm or downloading from Hugging Face.
 	@${MKDIR} ${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION}/tools/ui/dist && \
 		${CP} -r ${WRKDIR}/llama-${LLAMA_CPP_VERSION}/* \
 		${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION}/tools/ui/dist/
 
 post-patch:
 	# Apply llama.cpp laguna model patch (adds llama_model_laguna class)
 	#@${PATCH} -d ${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION} -p1 \
 	#	< ${WRKSRC}/llama/compat/models/003-llama-cpp-laguna.patch
 	# Apply Ollama llama.cpp compat hooks patch (vision model support)
 	#@${PATCH} -d ${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION} -p1 \
 	#	< ${WRKSRC}/llama/compat/001-llama-cpp-hooks.patch
 	# Fix llama-ui-embed crash when called without asset_dir argument
 	@${REINPLACE_CMD} -e 's|const std::string asset_dir = argv\[3\];|const std::string asset_dir = (argc > 3) ? argv[3] : "";|g' \
 		${WRKDIR}/llama.cpp-${LLAMA_CPP_VERSION}/tools/ui/embed.cpp
 	# FreeBSD: copy llm/llm_freebsd.go for LlamaServerSysProcAttr
 	@${CP} ${FILESDIR}/llm_freebsd.go ${WRKSRC}/llm/
 	# update version in version.go
 	@${REINPLACE_CMD} -e 's|var Version string = "0.0.0"|var Version string = "${PORTVERSION}"|g' \
 		${WRKSRC}/version/version.go
+	# expand WRKSRC in cmake/mlx/CMakeLists.txt
+	@${REINPLACE_CMD} -e 's|%%WRKSRC%%|${WRKSRC}|g' \
+		${WRKSRC}/cmake/mlx/CMakeLists.txt
 
 pre-build-CPU-on:
 	@${MKDIR} ${WRKSRC}/build && \
 		cd ${WRKSRC}/build && \
 		${CMAKE_BIN} ${_CMAKE_FLAGS} .. && \
 		${MAKE_CMD} ollama-llama-server-local
 
 pre-build-VULKAN-on:
 .if !${PORT_OPTIONS:MCPU} && !${PORT_OPTIONS:MMLX}
 	@${MKDIR} ${WRKSRC}/build && \
 		cd ${WRKSRC}/build && \
 		${CMAKE_BIN} ${_CMAKE_FLAGS} ..
 .endif
 	@cd ${WRKSRC}/build && \
 		${MAKE_CMD} ollama-llama-server-vulkan
 
 post-patch-MLX-on:
 	# FreeBSD compatibility fix: netinet/in.h (defines IPPROTO_TCP) is not
 	# pulled in transitively by netinet/tcp.h on FreeBSD as it is on Linux.
 	@${AWK} '/^#include <netinet\/tcp.h>/{print "#include <netinet/in.h>";print;next}1' \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/distributed/ring/ring.cpp > \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/distributed/ring/ring.cpp.new && \
 		${MV} ${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/distributed/ring/ring.cpp.new \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/distributed/ring/ring.cpp
 	# FreeBSD memory size fix: add hw.physmem sysctl query so MLX sets its
 	# memory_limit from actual RAM instead of using the 8 GB hardcoded fallback.
 	${INSTALL_DATA} ${FILESDIR}/freebsd_memory.h \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/backend/no_gpu/freebsd_memory.h
 	@${AWK} '/^#elif defined\(__linux__\)/{print "#elif defined(__FreeBSD__)";print "#include \"mlx/backend/no_gpu/freebsd_memory.h\"";print;next}1' \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/backend/no_gpu/allocator.cpp > \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/backend/no_gpu/allocator.cpp.new && \
 		${MV} ${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/backend/no_gpu/allocator.cpp.new \
 		${WRKDIR}/mlx-${MLX_CORE_VERSION}/mlx/backend/no_gpu/allocator.cpp
+	# FreeBSD fix for mlx-c: Use pre-downloaded mlx source via environment variable
+	@${AWK} '/if\(MLX_C_USE_SYSTEM_MLX\)/{print "# FreeBSD: Use pre-downloaded mlx source if OLLAMA_MLX_SOURCE is set";print "if(DEFINED ENV{OLLAMA_MLX_SOURCE})";print "  FetchContent_Declare(";print "    mlx";print "    SOURCE_DIR \"$$ENV{OLLAMA_MLX_SOURCE}\"";print "  )";print "  FetchContent_MakeAvailable(mlx)";print "elseif(MLX_C_USE_SYSTEM_MLX)";next}1' \
+		${WRKDIR}/mlx-c-${MLX_C_VERSION}/CMakeLists.txt > \
+		${WRKDIR}/mlx-c-${MLX_C_VERSION}/CMakeLists.txt.new && \
+		${MV} ${WRKDIR}/mlx-c-${MLX_C_VERSION}/CMakeLists.txt.new \
+		${WRKDIR}/mlx-c-${MLX_C_VERSION}/CMakeLists.txt
 
 pre-build-MLX-on:
 	@${MKDIR} ${WRKSRC}/build-mlx && \
 		cd ${WRKSRC}/build-mlx && \
 		OLLAMA_MLX_SOURCE=${WRKDIR}/mlx-${MLX_CORE_VERSION} \
 		OLLAMA_MLX_C_SOURCE=${WRKDIR}/mlx-c-${MLX_C_VERSION} \
 		${CMAKE_BIN} ${_CMAKE_FLAGS} \
 		-DFETCHCONTENT_FULLY_DISCONNECTED:BOOL=ON \
 		-DUSE_SYSTEM_FMT:BOOL=ON \
 		-DFETCHCONTENT_SOURCE_DIR_JSON:PATH=${WRKDIR}/json \
 		-DOLLAMA_SOURCE_DIR:PATH=${WRKSRC} \
 		-S ${WRKSRC}/cmake/mlx \
 		-B ${WRKSRC}/build-mlx
 	@cd ${WRKSRC}/build-mlx && \
 		${MAKE_CMD} mlx mlxc
 
 post-install: # pending https://github.com/ollama/ollama/issues/6407
 	${INSTALL_SCRIPT} ${FILESDIR}/ollama-limit-gpu-layers ${STAGEDIR}${PREFIX}/bin
 
 post-install-CPU-on:
 	@${MKDIR} ${STAGEDIR}${PREFIX}/lib/ollama
 	# versioned libggml-base
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libggml-base.so.${GGML_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${LN} -sf libggml-base.so.${GGML_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/libggml-base.so.0
 	${LN} -sf libggml-base.so.0 \
 		${STAGEDIR}${PREFIX}/lib/ollama/libggml-base.so
 	# versioned libggml
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libggml.so.${GGML_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${LN} -sf libggml.so.${GGML_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/libggml.so.0
 	${LN} -sf libggml.so.0 \
 		${STAGEDIR}${PREFIX}/lib/ollama/libggml.so
 	# versioned libllama-common
-	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama-common.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama-common.so.${LLAMA_CPP_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/
-	${LN} -sf libllama-common.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
-		${STAGEDIR}${PREFIX}/lib/ollama/libllama-common.so.0
-	${LN} -sf libllama-common.so.0 \
+	${LN} -sf libllama-common.so.${LLAMA_CPP_SO_VERSION} \
+		${STAGEDIR}${PREFIX}/lib/ollama/libllama-common.so.${LLAMA_CPP_SO_VERSION:R:R}
+	${LN} -sf libllama-common.so.${LLAMA_CPP_SO_VERSION:R:R} \
 		${STAGEDIR}${PREFIX}/lib/ollama/libllama-common.so
 	# versioned libllama
-	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama.so.${LLAMA_CPP_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/
-	${LN} -sf libllama.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
-		${STAGEDIR}${PREFIX}/lib/ollama/libllama.so.0
-	${LN} -sf libllama.so.0 \
+	${LN} -sf libllama.so.${LLAMA_CPP_SO_VERSION} \
+		${STAGEDIR}${PREFIX}/lib/ollama/libllama.so.${LLAMA_CPP_SO_VERSION:R:R}
+	${LN} -sf libllama.so.${LLAMA_CPP_SO_VERSION:R:R} \
 		${STAGEDIR}${PREFIX}/lib/ollama/libllama.so
 	# versioned libmtmd
-	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libmtmd.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libmtmd.so.${LLAMA_CPP_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/
-	${LN} -sf libmtmd.so.0.0.${LLAMA_CPP_VERSION:S/b//} \
+	${LN} -sf libmtmd.so.${LLAMA_CPP_SO_VERSION} \
 		${STAGEDIR}${PREFIX}/lib/ollama/libmtmd.so.0
 	${LN} -sf libmtmd.so.0 \
 		${STAGEDIR}${PREFIX}/lib/ollama/libmtmd.so
 	# unversioned impl libs needed at runtime by llama-server and llama-quantize
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama-server-impl.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libllama-quantize-impl.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	# binaries
 	${INSTALL_PROGRAM} ${WRKSRC}/build/lib/ollama/llama-server \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${INSTALL_PROGRAM} ${WRKSRC}/build/lib/ollama/llama-quantize \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	# cpu-variant SIMD backend DSOs
 	@for f in ${WRKSRC}/build/lib/ollama/libggml-cpu*.so; do \
 		${INSTALL_LIB} $$f ${STAGEDIR}${PREFIX}/lib/ollama/; \
 	done
 
 post-install-VULKAN-on:
 	@${MKDIR} ${STAGEDIR}${PREFIX}/lib/ollama/vulkan
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/vulkan/libggml-vulkan.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/vulkan/
 	# Fix RPATH: libggml-base.so.0 lives in the parent directory
 	patchelf --set-rpath '$$ORIGIN:$$ORIGIN/..' \
 		${STAGEDIR}${PREFIX}/lib/ollama/vulkan/libggml-vulkan.so
 
 post-install-MLX-on:
 	@${MKDIR} ${STAGEDIR}${PREFIX}/lib/ollama
 	${INSTALL_LIB} ${WRKSRC}/build-mlx/lib/ollama/libmlx.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${INSTALL_LIB} ${WRKSRC}/build-mlx/lib/ollama/libmlxc.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	patchelf --set-rpath '$$ORIGIN' ${STAGEDIR}${PREFIX}/lib/ollama/libmlx.so
 	patchelf --set-rpath '$$ORIGIN' ${STAGEDIR}${PREFIX}/lib/ollama/libmlxc.so
 
 do-test:
 	@cd ${WRKSRC} && \
 		${SETENVI} ${WRK_ENV} ${MAKE_ENV} ${GO_ENV} ${GO_CMD} test ./...
 
 .include <bsd.port.mk>
diff --git a/misc/ollama/distinfo b/misc/ollama/distinfo
index 8903edd52625..9219340170f0 100644
--- a/misc/ollama/distinfo
+++ b/misc/ollama/distinfo
@@ -1,15 +1,17 @@
-TIMESTAMP = 1785285890
-SHA256 (go/misc_ollama/ollama-v0.32.5/b10091.tar.gz) = 37da9cfee7ae5b0c9ea4e712e0d21c8242e5e7a781c02035498efb8a21a7e329
-SIZE (go/misc_ollama/ollama-v0.32.5/b10091.tar.gz) = 35959727
-SHA256 (go/misc_ollama/ollama-v0.32.5/llama-b10091-ui.tar.gz) = 5690c3f81031adfd1f229a954506cb9017a9c281ead28c2bf450d7fdbfab3cc3
-SIZE (go/misc_ollama/ollama-v0.32.5/llama-b10091-ui.tar.gz) = 2872825
-SHA256 (go/misc_ollama/ollama-v0.32.5/v0.31.2.tar.gz) = bdb9b619f80962dd00c0bffb65e59c53f565c2b550f189a1467f8bc6089401ab
-SIZE (go/misc_ollama/ollama-v0.32.5/v0.31.2.tar.gz) = 4251596
-SHA256 (go/misc_ollama/ollama-v0.32.5/fba4470b89073180056c9ea46c443051375f7399.tar.gz) = 398d3d4c6458679dfc9061a3ba56101358279790bc038f60d1f534eac420d552
-SIZE (go/misc_ollama/ollama-v0.32.5/fba4470b89073180056c9ea46c443051375f7399.tar.gz) = 174039
-SHA256 (go/misc_ollama/ollama-v0.32.5/json.tar.xz) = d6c65aca6b1ed68e7a182f4757257b107ae403032760ed6ef121c9d55e81757d
-SIZE (go/misc_ollama/ollama-v0.32.5/json.tar.xz) = 110988
-SHA256 (go/misc_ollama/ollama-v0.32.5/v0.32.5.mod) = a67e90c190ca0c106b7db76216a29c0ec4cab2ad1bfbde49cdced90fe78193bd
-SIZE (go/misc_ollama/ollama-v0.32.5/v0.32.5.mod) = 4622
-SHA256 (go/misc_ollama/ollama-v0.32.5/v0.32.5.zip) = adb86caaee759e69dfd416f6998a65a046929f822ec189962b4a8319efafb7c7
-SIZE (go/misc_ollama/ollama-v0.32.5/v0.32.5.zip) = 30371899
+TIMESTAMP = 1789036710
+SHA256 (go/misc_ollama/ollama-v0.34.0/b10760.tar.gz) = 8986285d34f7531e84b1ee6d35c88989fc6f3cf895f60610594997b8b7bad443
+SIZE (go/misc_ollama/ollama-v0.34.0/b10760.tar.gz) = 37131759
+SHA256 (go/misc_ollama/ollama-v0.34.0/llama-b10760-ui.tar.gz) = b7f7d173c0faa6582896d2d23603c1bb2d963eedf8845db0a8399d2c1e0ce360
+SIZE (go/misc_ollama/ollama-v0.34.0/llama-b10760-ui.tar.gz) = 3081120
+SHA256 (go/misc_ollama/ollama-v0.34.0/v0.31.2.tar.gz) = bdb9b619f80962dd00c0bffb65e59c53f565c2b550f189a1467f8bc6089401ab
+SIZE (go/misc_ollama/ollama-v0.34.0/v0.31.2.tar.gz) = 4251596
+SHA256 (go/misc_ollama/ollama-v0.34.0/fba4470b89073180056c9ea46c443051375f7399.tar.gz) = 398d3d4c6458679dfc9061a3ba56101358279790bc038f60d1f534eac420d552
+SIZE (go/misc_ollama/ollama-v0.34.0/fba4470b89073180056c9ea46c443051375f7399.tar.gz) = 174039
+SHA256 (go/misc_ollama/ollama-v0.34.0/json.tar.xz) = d6c65aca6b1ed68e7a182f4757257b107ae403032760ed6ef121c9d55e81757d
+SIZE (go/misc_ollama/ollama-v0.34.0/json.tar.xz) = 110988
+SHA256 (go/misc_ollama/ollama-v0.34.0/v0.34.0.mod) = b70e409c579c9184be085ae9ab2e06b02e2426fb6781548580596ba341105527
+SIZE (go/misc_ollama/ollama-v0.34.0/v0.34.0.mod) = 4640
+SHA256 (go/misc_ollama/ollama-v0.34.0/v0.34.0.zip) = 3ffdf867ded14b264a96526e6ea86be4f4ff0e5ceca812428d7ab9e7d461e32e
+SIZE (go/misc_ollama/ollama-v0.34.0/v0.34.0.zip) = 31535194
+SHA256 (go/misc_ollama/ollama-v0.34.0/mlc-ai-xgrammar-v0.2.5_GH0.tar.gz) = 7cf5e69ba0bc2695a9f1a97eeb4cc567aac3d644fd329e7d8072ab6f3517b777
+SIZE (go/misc_ollama/ollama-v0.34.0/mlc-ai-xgrammar-v0.2.5_GH0.tar.gz) = 1071195
diff --git a/misc/ollama/files/patch-CMakeLists.txt b/misc/ollama/files/patch-CMakeLists.txt
deleted file mode 100644
index 0d4b1cc5a0a6..000000000000
--- a/misc/ollama/files/patch-CMakeLists.txt
+++ /dev/null
@@ -1,13 +0,0 @@
---- CMakeLists.txt.orig	1979-11-30 08:00:00 UTC
-+++ CMakeLists.txt
-@@ -45,6 +45,10 @@ if(APPLE)
-     set(CMAKE_BUILD_RPATH "@loader_path")
-     set(CMAKE_INSTALL_RPATH "@loader_path")
-     set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
-+elseif(UNIX)
-+    set(CMAKE_BUILD_RPATH "$ORIGIN")
-+    set(CMAKE_INSTALL_RPATH "$ORIGIN")
-+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
- endif()
- 
- set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
diff --git a/misc/ollama/files/patch-cmake_mlx_CMakeLists.txt b/misc/ollama/files/patch-cmake_mlx_CMakeLists.txt
new file mode 100644
index 000000000000..7ae896a0db9e
--- /dev/null
+++ b/misc/ollama/files/patch-cmake_mlx_CMakeLists.txt
@@ -0,0 +1,27 @@
+--- cmake/mlx/CMakeLists.txt.orig	1979-11-30 08:00:00 UTC
++++ cmake/mlx/CMakeLists.txt
+@@ -63,14 +63,16 @@ set(XGRAMMAR_VERSION v0.2.5)
+ 
+ include(FetchContent)
+ set(XGRAMMAR_VERSION v0.2.5)
+-FetchContent_Declare(xgrammar
+-    GIT_REPOSITORY "https://github.com/mlc-ai/xgrammar.git"
+-    GIT_TAG ${XGRAMMAR_VERSION}
+-    GIT_SHALLOW TRUE
+-    GIT_SUBMODULES 3rdparty/dlpack
+-    # Do not add XGrammar's Python-oriented CMake project.
+-    SOURCE_SUBDIR cmake/ollama)
+-FetchContent_MakeAvailable(xgrammar)
++#FetchContent_Declare(xgrammar
++#    GIT_REPOSITORY "https://github.com/mlc-ai/xgrammar.git"
++#    GIT_TAG ${XGRAMMAR_VERSION}
++#    GIT_SHALLOW TRUE
++#    GIT_SUBMODULES 3rdparty/dlpack
++#    # Do not add XGrammar's Python-oriented CMake project.
++#    SOURCE_SUBDIR cmake/ollama)
++#FetchContent_MakeAvailable(xgrammar)
++
++set(xgrammar_SOURCE_DIR "%%WRKSRC%%/xgrammar")
+ 
+ file(GLOB_RECURSE XGRAMMAR_SOURCES CONFIGURE_DEPENDS "${xgrammar_SOURCE_DIR}/cpp/*.cc")
+ list(FILTER XGRAMMAR_SOURCES EXCLUDE REGEX "/cpp/tvm_ffi/.*\\.cc$")
diff --git a/misc/ollama/files/patch-x_imagegen_memory.go b/misc/ollama/files/patch-x_imagegen_memory.go
deleted file mode 100644
index 86279bc6592e..000000000000
--- a/misc/ollama/files/patch-x_imagegen_memory.go
+++ /dev/null
@@ -1,11 +0,0 @@
---- x/imagegen/memory.go.orig	2026-03-12 23:02:15 UTC
-+++ x/imagegen/memory.go
-@@ -31,7 +31,7 @@ func CheckPlatformSupport() error {
- 			return fmt.Errorf("image generation on macOS requires Apple Silicon (arm64), got %s", runtime.GOARCH)
- 		}
- 		return nil
--	case "linux", "windows":
-+	case "linux", "windows", "freebsd":
- 		// Linux/Windows: CUDA support (requires mlx or cuda build)
- 		// The actual backend availability is checked at runtime
- 		return nil
diff --git a/misc/ollama/files/patch-x_imagegen_mlx_mlx.go b/misc/ollama/files/patch-x_imagegen_mlx_mlx.go
deleted file mode 100644
index 426c4529f9b1..000000000000
--- a/misc/ollama/files/patch-x_imagegen_mlx_mlx.go
+++ /dev/null
@@ -1,50 +0,0 @@
--- FreeBSD compatibility fixes for MLX image generation CGO code.
--- 1. Add FreeBSD LDFLAGS (-lc++ -ldl) required for C++ runtime and dynamic loading.
--- 2. Fall back to CPU stream in default_stream() when GPU stream is unavailable (no Metal on FreeBSD).
--- 3. Use CPU stream for safetensors loading on FreeBSD (Metal eval_gpu not implemented).
--- 4. Wrap mlx_load_safetensors with safe init mode to capture C++ exceptions as Go errors
---    instead of process exit(1), enabling proper error reporting to the user.
---- x/imagegen/mlx/mlx.go.orig	2026-06-03 07:41:51 UTC
-+++ x/imagegen/mlx/mlx.go
-@@ -4,6 +4,7 @@ package mlx
- #cgo CFLAGS: -O3 -I${SRCDIR}/../../mlxrunner/mlx/include -I${SRCDIR}
- #cgo darwin LDFLAGS: -lc++ -framework Metal -framework Foundation -framework Accelerate
- #cgo linux LDFLAGS: -lstdc++ -ldl
-+#cgo freebsd LDFLAGS: -lc++ -ldl
- #cgo windows LDFLAGS: -lstdc++
- 
- // Use generated wrappers instead of direct MLX headers
-@@ -23,6 +24,9 @@ static inline mlx_stream default_stream() {
- static inline mlx_stream default_stream() {
-     if (_default_stream.ctx == NULL) {
-         _default_stream = mlx_default_gpu_stream_new();
-+        if (_default_stream.ctx == NULL) {
-+            _default_stream = mlx_default_cpu_stream_new();
-+        }
-     }
-     return _default_stream;
- }
-@@ -1512,13 +1516,21 @@ func LoadSafetensorsNative(path string) (*SafetensorsF
- 	defer C.free(unsafe.Pointer(cPath))
- 
- 	stream := C.default_stream()
--	if runtime.GOOS == "darwin" {
-+	if runtime.GOOS == "darwin" || runtime.GOOS == "freebsd" {
- 		stream = C.cpu_stream()
- 	}
- 
- 	var arrays C.mlx_map_string_to_array
- 	var metadata C.mlx_map_string_to_string
--	if C.mlx_load_safetensors(&arrays, &metadata, cPath, stream) != 0 {
-+	C.mlx_set_safe_init_mode()
-+	ret := C.mlx_load_safetensors(&arrays, &metadata, cPath, stream)
-+	if C.mlx_had_init_error() != 0 {
-+		msg := C.GoString(C.mlx_get_init_error())
-+		C.mlx_set_default_error_mode()
-+		return nil, fmt.Errorf("failed to load safetensors: %s (mlx error: %s)", path, msg)
-+	}
-+	C.mlx_set_default_error_mode()
-+	if ret != 0 {
- 		return nil, fmt.Errorf("failed to load safetensors: %s", path)
- 	}
- 	return &SafetensorsFile{arrays: arrays, metadata: metadata}, nil
diff --git a/misc/ollama/files/patch-x_imagegen_models_zimage_vae.go b/misc/ollama/files/patch-x_imagegen_models_zimage_vae.go
deleted file mode 100644
index 1cd9e0eaa6b8..000000000000
--- a/misc/ollama/files/patch-x_imagegen_models_zimage_vae.go
+++ /dev/null
@@ -1,35 +0,0 @@
---- x/imagegen/models/zimage/vae.go.orig	1979-11-30 08:00:00 UTC
-+++ x/imagegen/models/zimage/vae.go
-@@ -332,6 +332,16 @@ func (rb *ResnetBlock2D) Forward(x *mlx.Array) *mlx.Ar
- 
- // Forward applies the ResNet block with staged evaluation
- func (rb *ResnetBlock2D) Forward(x *mlx.Array) *mlx.Array {
-+	// Keep x alive across intermediate Eval calls (cleanup() would free it otherwise).
-+	// The residual connection at the end needs the original x.
-+	wasKept := x.Kept()
-+	mlx.Keep(x)
-+	defer func() {
-+		if !wasKept {
-+			x.Free()
-+		}
-+	}()
-+
- 	var h *mlx.Array
- 
- 	// Stage 1: norm1
-@@ -461,6 +471,15 @@ func (ab *VAEAttentionBlock) Forward(x *mlx.Array) *ml
- // Input and output are in NHWC format [B, H, W, C]
- func (ab *VAEAttentionBlock) Forward(x *mlx.Array) *mlx.Array {
- 	residual := x
-+	// Keep residual alive across intermediate Eval calls.
-+	// The residual addition at stage 3 needs the original input.
-+	wasKept := residual.Kept()
-+	mlx.Keep(residual)
-+	defer func() {
-+		if !wasKept {
-+			residual.Free()
-+		}
-+	}()
- 	shape := x.Shape()
- 	B := shape[0]
- 	H := shape[1]
diff --git a/misc/ollama/files/patch-x_imagegen_server.go b/misc/ollama/files/patch-x_imagegen_server.go
deleted file mode 100644
index c9169dca35df..000000000000
--- a/misc/ollama/files/patch-x_imagegen_server.go
+++ /dev/null
@@ -1,13 +0,0 @@
---- x/imagegen/server.go.orig	2026-04-23 17:32:09 UTC
-+++ x/imagegen/server.go
-@@ -55,7 +55,9 @@ func NewServer(modelName string) (*Server, error) {
- 	return &Server{
- 		modelName: modelName,
- 		done:      make(chan error, 1),
--		client:    &http.Client{Timeout: 10 * time.Minute},
-+		// No client-level timeout: image generation on CPU can take many minutes.
-+		// Cancellation is handled via request context.
-+		client: &http.Client{},
- 	}, nil
- }
- 
diff --git a/misc/ollama/files/patch-x_mlxrunner_client.go b/misc/ollama/files/patch-x_mlxrunner_client.go
index 042b1d58222b..ae43bbc1c58a 100644
--- a/misc/ollama/files/patch-x_mlxrunner_client.go
+++ b/misc/ollama/files/patch-x_mlxrunner_client.go
@@ -1,11 +1,11 @@
---- x/mlxrunner/client.go.orig	2026-04-18 04:07:37 UTC
+--- x/mlxrunner/client.go.orig	1979-11-30 08:00:00 UTC
 +++ x/mlxrunner/client.go
-@@ -366,6 +366,8 @@ func (c *Client) Load(ctx context.Context, _ ml.System
+@@ -353,6 +353,8 @@ func (c *Client) Load(ctx context.Context, _ ml.System
  	switch runtime.GOOS {
  	case "linux":
  		libPathEnvVar = "LD_LIBRARY_PATH"
 +	case "freebsd":
 +		libPathEnvVar = "LD_LIBRARY_PATH"
  	case "windows":
  		libPathEnvVar = "PATH"
  	}
diff --git a/misc/ollama/files/patch-x_mlxrunner_mlx_CMakeLists.txt b/misc/ollama/files/patch-x_mlxrunner_mlx_CMakeLists.txt
new file mode 100644
index 000000000000..2b408deec400
--- /dev/null
+++ b/misc/ollama/files/patch-x_mlxrunner_mlx_CMakeLists.txt
@@ -0,0 +1,30 @@
+-- FreeBSD: Use pre-downloaded mlx-c source directory instead of FetchContent from Git.
+-- The port downloads mlx-c as a distfile and extracts it to WRKDIR, then passes
+-- its location via OLLAMA_MLX_C_SOURCE environment variable to the CMake build.
+--- x/mlxrunner/mlx/CMakeLists.txt.orig	1979-11-30 08:00:00 UTC
++++ x/mlxrunner/mlx/CMakeLists.txt
+@@ -23,11 +23,19 @@ string(STRIP "${MLX_C_GIT_TAG}" MLX_C_GIT_TAG)
+ file(READ "${CMAKE_CURRENT_LIST_DIR}/../../../MLX_C_VERSION" MLX_C_GIT_TAG)
+ string(STRIP "${MLX_C_GIT_TAG}" MLX_C_GIT_TAG)
+ 
+-FetchContent_Declare(
+-  mlx-c
+-  GIT_REPOSITORY "https://github.com/ml-explore/mlx-c.git"
+-  GIT_TAG ${MLX_C_GIT_TAG}
+-)
++# FreeBSD: Use pre-downloaded mlx-c source if OLLAMA_MLX_C_SOURCE is set
++if(DEFINED ENV{OLLAMA_MLX_C_SOURCE})
++  FetchContent_Declare(
++    mlx-c
++    SOURCE_DIR "$ENV{OLLAMA_MLX_C_SOURCE}"
++  )
++else()
++  FetchContent_Declare(
++    mlx-c
++    GIT_REPOSITORY "https://github.com/ml-explore/mlx-c.git"
++    GIT_TAG ${MLX_C_GIT_TAG}
++  )
++endif()
+ 
+ FetchContent_MakeAvailable(mlx-c)
+ 
diff --git a/misc/ollama/files/patch-x_mlxrunner_mlx_dynamic.go b/misc/ollama/files/patch-x_mlxrunner_mlx_dynamic.go
index 76e6d0bec723..7ca8c2caff6c 100644
--- a/misc/ollama/files/patch-x_mlxrunner_mlx_dynamic.go
+++ b/misc/ollama/files/patch-x_mlxrunner_mlx_dynamic.go
@@ -1,29 +1,29 @@
---- x/mlxrunner/mlx/dynamic.go.orig	2026-04-09 01:42:19 UTC
+--- x/mlxrunner/mlx/dynamic.go.orig	1979-11-30 08:00:00 UTC
 +++ x/mlxrunner/mlx/dynamic.go
-@@ -83,7 +83,7 @@ func libOllamaRoots() []string {
- 		case "darwin":
+@@ -98,7 +98,7 @@ func libOllamaRoots() []string {
  			roots = append(roots, filepath.Join(exeDir, "lib", "ollama"))
+ 			roots = append(roots, filepath.Join(exeDir, "..", "lib", "ollama"))
  			roots = append(roots, exeDir) // app bundle: Contents/Resources/
 -		case "linux":
 +		case "linux", "freebsd":
  			roots = append(roots, filepath.Join(exeDir, "..", "lib", "ollama"))
  		case "windows":
  			roots = append(roots, filepath.Join(exeDir, "lib", "ollama"))
-@@ -143,7 +143,7 @@ func prependLibraryPath(dir string) {
+@@ -175,7 +175,7 @@ func prependLibraryPath(dir string) {
  	switch runtime.GOOS {
  	case "darwin":
  		envVar = "DYLD_LIBRARY_PATH"
 -	case "linux":
 +	case "linux", "freebsd":
  		envVar = "LD_LIBRARY_PATH"
  	default:
  		return
-@@ -157,7 +157,7 @@ func init() {
+@@ -189,7 +189,7 @@ func init() {
  
  func init() {
  	switch runtime.GOOS {
 -	case "darwin", "linux", "windows":
 +	case "darwin", "linux", "freebsd", "windows":
  	default:
  		return
  	}