diff --git a/misc/ollama/Makefile b/misc/ollama/Makefile
index f10ae5724905..e78c28f88d48 100644
--- a/misc/ollama/Makefile
+++ b/misc/ollama/Makefile
@@ -1,162 +1,162 @@
 PORTNAME=	ollama
 DISTVERSIONPREFIX=	v
 DISTVERSION=	0.18.3
-PORTREVISION=	1
+PORTREVISION=	2
 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
 
 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 \
 		patchelf:sysutils/patchelf
 
 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}"
 
 MLX_CORE_VERSION=	0.31.1
 MLX_C_VERSION=		0.6.0
 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 MLX
 
 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=	${LOCALBASE}/lib/cmake/fmt/fmt-config.cmake:devel/libfmt
 MLX_LIB_DEPENDS=	libopenblas.so:math/openblas
 MLX_PLIST_FILES=	lib/ollama/libmlx.so \
 			lib/ollama/libmlxc.so
 _CMAKE_FLAGS=	-DCMAKE_BUILD_TYPE=Release -DGGML_BACKEND_DL=ON -DGGML_BACKEND_DIR=${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: 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
 
 pre-build-MLX-on:
 	@${MKDIR} ${WRKSRC}/build && \
 		cd ${WRKSRC}/build && \
 		OLLAMA_MLX_SOURCE=${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/
 	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/files/patch-x_imagegen_models_zimage_vae.go b/misc/ollama/files/patch-x_imagegen_models_zimage_vae.go
new file mode 100644
index 000000000000..0c9e990b52ba
--- /dev/null
+++ b/misc/ollama/files/patch-x_imagegen_models_zimage_vae.go
@@ -0,0 +1,35 @@
+--- x/imagegen/models/zimage/vae.go.orig	1979-11-30 00:00:00.000000000 -0800
++++ x/imagegen/models/zimage/vae.go
+@@ -332,6 +332,16 @@
+ 
+ // 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 @@
+ // 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
new file mode 100644
index 000000000000..a49851568dec
--- /dev/null
+++ b/misc/ollama/files/patch-x_imagegen_server.go
@@ -0,0 +1,26 @@
+-- same issue as first chunk: https://github.com/ollama/ollama/issues/15107
+
+--- x/imagegen/server.go.orig
++++ x/imagegen/server.go
+@@ -55,7 +55,9 @@
+ 	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
+ }
+ 
+@@ -116,8 +118,8 @@
+ 	cmd := exec.Command(exe, "runner", "--imagegen-engine", "--model", s.modelName, "--port", strconv.Itoa(port))
+ 	cmd.Env = os.Environ()
+ 
+-	// On Linux, set LD_LIBRARY_PATH to include MLX library directories
+-	if runtime.GOOS == "linux" {
++	// On Linux and FreeBSD, set LD_LIBRARY_PATH to include MLX library directories
++	if runtime.GOOS == "linux" || runtime.GOOS == "freebsd" {
+ 		// Build library paths: start with LibOllamaPath, then add any mlx_* subdirectories
+ 		libraryPaths := []string{ml.LibOllamaPath}
+ 		if mlxDirs, err := filepath.Glob(filepath.Join(ml.LibOllamaPath, "mlx_*")); err == nil {