diff --git a/misc/ollama/Makefile b/misc/ollama/Makefile
index 2f0c7ddf8920..1a3552c87d85 100644
--- a/misc/ollama/Makefile
+++ b/misc/ollama/Makefile
@@ -1,128 +1,150 @@
 PORTNAME=	ollama
 DISTVERSIONPREFIX=	v
 DISTVERSION=	0.18.2
+PORTREVISION=	1
 CATEGORIES=	misc # machine-learning
 
 MAINTAINER=	yuri@FreeBSD.org
 COMMENT=	Run Llama 2, Mistral, and other large language models
 WWW=		https://ollama.com \
 		https://github.com/ollama/ollama
 
-BROKEN_i386=	fails to compile: x/mlxrunner/mlx/memory.go:40:11: 1 << (4 * 10) (untyped int constant 1099511627776) overflows int
-
 LICENSE=	MIT
 LICENSE_FILE=	${WRKSRC}/LICENSE
 
+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
 
 USES=		cmake:indirect go:1.24+,modules localbase pkgconfig
 
 GO_MODULE=	github.com/yurivict/${PORTNAME} # fork with FreeBSD patches
 GO_TARGET=	.
 GO_ENV+=	CGO_CXXFLAGS="${CXXFLAGS}"
 
-.include <bsd.port.options.mk>
-
-.if ${PORT_OPTIONS:MMLX}
-GO_BUILDFLAGS+=	-tags mlx
-.endif
+MLX_CORE_VERSION=	0.30.6
+MLX_C_VERSION=		0.5.0
+JSON_VERSION=		3.11.3
 
 PLIST_FILES=	bin/${PORTNAME} \
 		bin/ollama-limit-gpu-layers
 
 OPTIONS_GROUP=		BACKENDS
 OPTIONS_GROUP_BACKENDS=	CPU VULKAN MLX
-OPTIONS_DEFAULT=	CPU VULKAN
+OPTIONS_DEFAULT=	CPU VULKAN MLX
 
-CPU_DESC=		Build CPU backend shared libraries
+CPU_DESC=		Build CPU backend shared libraries for various SIMD instruction sets
 CPU_PLIST_FILES=	lib/ollama/libggml-base.so \
 			lib/ollama/libggml-cpu-alderlake.so \
 			lib/ollama/libggml-cpu-haswell.so \
 			lib/ollama/libggml-cpu-icelake.so \
 			lib/ollama/libggml-cpu-sandybridge.so \
 			lib/ollama/libggml-cpu-skylakex.so \
 			lib/ollama/libggml-cpu-sse42.so \
 			lib/ollama/libggml-cpu-x64.so
 
 VULKAN_DESC=		Build Vulkan GPU backend shared library
 VULKAN_BUILD_DEPENDS=	glslc:graphics/shaderc \
 			${LOCALBASE}/include/vulkan/vulkan.h:graphics/vulkan-headers
 VULKAN_LIB_DEPENDS=	libvulkan.so:graphics/vulkan-loader
 VULKAN_PLIST_FILES=	lib/ollama/libggml-vulkan.so
 
 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
-MLX_CMAKE_ON=		-DMLX_ENGINE:BOOL=ON
 
 _CMAKE_FLAGS=	-DCMAKE_BUILD_TYPE=Release -DGGML_BACKEND_DL=ON -DGGML_BACKEND_DIR=${PREFIX}/lib/ollama
 
+.include <bsd.port.options.mk>
+
+.if ${PORT_OPTIONS:MMLX}
+GO_BUILDFLAGS+=	-tags mlx
+DISTFILES+=	v${MLX_CORE_VERSION}.tar.gz:mlxsrc \
+		v${MLX_C_VERSION}.tar.gz:mlxcsrc \
+		json.tar.xz:jsonsrc
+MASTER_SITES+=	https://github.com/ml-explore/mlx/archive/refs/tags/:mlxsrc \
+		https://github.com/ml-explore/mlx-c/archive/refs/tags/:mlxcsrc \
+		https://github.com/nlohmann/json/releases/download/v${JSON_VERSION}/:jsonsrc
+.endif
+
 post-patch:
 	# change import path to the fork
 	@cd ${WRKSRC} && \
 		(${GREP} -rl ollama/ollama | ${XARGS} ${REINPLACE_CMD} -i '' -e 's|ollama/ollama|yurivict/ollama|g')
+	# update version in go.mod and version.go
 	@${REINPLACE_CMD} -e 's|var Version string = "0.0.0"|var Version string = "${PORTVERSION}"|g' \
 		${WRKSRC}/go.mod ${WRKSRC}/version/version.go
 
 pre-build-CPU-on:
 	@${MKDIR} ${WRKSRC}/build && \
 		cd ${WRKSRC}/build && \
 		${CMAKE_BIN} ${_CMAKE_FLAGS} .. && \
 		${MAKE_CMD} ggml-base && \
 		${MAKE_CMD} ggml-cpu
 
 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} ggml-vulkan
 
+post-patch-MLX-on:
+	# FreeBSD compatibility fix: add missing netinet/in.h include
+	@${AWK} '/^#include <fcntl.h>/{print;print "#include <netinet/in.h>";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
+
 pre-build-MLX-on:
-.if !${PORT_OPTIONS:MCPU} && !${PORT_OPTIONS:MVULKAN}
 	@${MKDIR} ${WRKSRC}/build && \
 		cd ${WRKSRC}/build && \
-		${CMAKE_BIN} ${_CMAKE_FLAGS} -DMLX_ENGINE:BOOL=ON ..
-.endif
-	# Patch MLX after it's fetched by CMake (FreeBSD compatibility fix)
-	@if [ -f ${WRKSRC}/build/_deps/mlx-src/mlx/distributed/ring/ring.cpp ]; then \
-		${REINPLACE_CMD} -i '' -e '/^#include <fcntl.h>/a\\
-#include <netinet/in.h>' \
-			${WRKSRC}/build/_deps/mlx-src/mlx/distributed/ring/ring.cpp; \
-	fi
+		OLLAMA_MLX_SOURCE=${WRKDIR}/mlx-${MLX_CORE_VERSION} \
+		OLLAMA_MLX_C_SOURCE=${WRKDIR}/mlx-c-${MLX_C_VERSION} \
+		${CMAKE_BIN} ${_CMAKE_FLAGS} \
+		-DMLX_ENGINE:BOOL=ON \
+		-DFETCHCONTENT_FULLY_DISCONNECTED:BOOL=ON \
+		-DFETCHCONTENT_SOURCE_DIR_JSON:PATH=${WRKDIR}/json \
+		-DUSE_SYSTEM_FMT:BOOL=ON \
+		-DMLX_BUILD_TESTS:BOOL=OFF \
+		-DMLX_BUILD_EXAMPLES:BOOL=OFF \
+		..
 	@cd ${WRKSRC}/build && \
 		${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
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libggml-base.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	@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
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libggml-vulkan.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 
 post-install-MLX-on:
 	@${MKDIR} ${STAGEDIR}${PREFIX}/lib/ollama
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libmlx.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 	${INSTALL_LIB} ${WRKSRC}/build/lib/ollama/libmlxc.so \
 		${STAGEDIR}${PREFIX}/lib/ollama/
 
 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 fbdc86e10cc6..e709a9fc964b 100644
--- a/misc/ollama/distinfo
+++ b/misc/ollama/distinfo
@@ -1,5 +1,11 @@
 TIMESTAMP = 1774149327
 SHA256 (go/misc_ollama/ollama-v0.18.2/v0.18.2.mod) = e660ffa37ad33611a4c1248967204e578f07ada9941e63ffe6a6fe3e95ac5039
 SIZE (go/misc_ollama/ollama-v0.18.2/v0.18.2.mod) = 4757
 SHA256 (go/misc_ollama/ollama-v0.18.2/v0.18.2.zip) = 2ad77971884d45d35ba80db3d610cdfb93e7a2f390ea68fe1745085ebf73da4a
 SIZE (go/misc_ollama/ollama-v0.18.2/v0.18.2.zip) = 23468279
+SHA256 (go/misc_ollama/ollama-v0.18.2/v0.30.6.tar.gz) = 04d8e3c5d3c3ae748145ffe51d3ad8b72493e1e8597fc0f9dc631ed620b12618
+SIZE (go/misc_ollama/ollama-v0.18.2/v0.30.6.tar.gz) = 4138686
+SHA256 (go/misc_ollama/ollama-v0.18.2/v0.5.0.tar.gz) = dcfc404d7004e6da70170c669dbc920913cb25a59c9f7dac781caf92e524cc86
+SIZE (go/misc_ollama/ollama-v0.18.2/v0.5.0.tar.gz) = 169957
+SHA256 (go/misc_ollama/ollama-v0.18.2/json.tar.xz) = d6c65aca6b1ed68e7a182f4757257b107ae403032760ed6ef121c9d55e81757d
+SIZE (go/misc_ollama/ollama-v0.18.2/json.tar.xz) = 110988
diff --git a/misc/ollama/pkg-message b/misc/ollama/pkg-message
index 4d6db4b1a43a..b286f8cf71d1 100644
--- a/misc/ollama/pkg-message
+++ b/misc/ollama/pkg-message
@@ -1,33 +1,33 @@
 [
 { type: install
   message: <<EOM
 You installed ollama: the AI model runner.
 
 To run ollama, plese open 2 terminals.
 1. In the first terminal, please run:
-   $ OLLAMA_NUM_PARALLEL=1 OLLAMA_DEBUG=1 LLAMA_DEBUG=1 ollama start
+   $ ollama serve
 2. In the second terminal, please run:
    $ ollama run gemma3
    or
    $ ollama run mistral
 
 This will download and run the specified AI model.
 You will be able to interact with it in plain English.
 
 Please see https://ollama.com/library for the list
 of all supported models.
 
 The command "ollama list" lists all models downloaded
 into your system.
 
 When the model fails to load into your GPU, please use
 the provided ollama-limit-gpu-layers script to create
 model flavors with different num_gpu parameters.
 
 ollama uses many gigabytes of disk space in your home directory,
 because advanced AI models are often very large.
 Please symlink ~/.ollama to a large disk if needed.
 
 EOM
 }
 ]