diff --git a/misc/llama-cpp/Makefile b/misc/llama-cpp/Makefile
index b0d90427a3ff..19a4a9357d69 100644
--- a/misc/llama-cpp/Makefile
+++ b/misc/llama-cpp/Makefile
@@ -1,45 +1,48 @@
 PORTNAME=	llama-cpp
 DISTVERSIONPREFIX=	b
-DISTVERSION=	3631
+DISTVERSION=	3647
 CATEGORIES=	misc # machine-learning
 
+PATCH_SITES=	https://github.com/${GH_ACCOUNT}/${GH_PROJECT}/commit/
+PATCHFILES=	121f915a09c1117d34aff6e8faf6d252aaf11027.patch:-p1 # Add missing pthread includes: https://github.com/ggerganov/llama.cpp/pull/9258
+
 MAINTAINER=	yuri@FreeBSD.org
 COMMENT=	Facebook's LLaMA model in C/C++ # '
 WWW=		https://github.com/ggerganov/llama.cpp
 
 LICENSE=	MIT
 LICENSE_FILE=	${WRKSRC}/LICENSE
 
 BROKEN_armv7=	clang crashes, see https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=278810
 
 USES=		cmake:testing compiler:c++11-lang python:run shebangfix
 USE_LDCONFIG=	yes
 
 USE_GITHUB=	yes
 GH_ACCOUNT=	ggerganov
 GH_PROJECT=	llama.cpp
 GH_TUPLE=	nomic-ai:kompute:4565194:kompute/kompute
 
 SHEBANG_GLOB=	*.py
 
 CMAKE_ON=	BUILD_SHARED_LIBS
 CMAKE_OFF=	LLAMA_BUILD_TESTS
 CMAKE_TESTING_ON=	LLAMA_BUILD_TESTS
 
 OPTIONS_DEFINE=		EXAMPLES VULKAN
 OPTIONS_DEFAULT=	VULKAN
 OPTIONS_SUB=		yes
 
 EXAMPLES_CMAKE_BOOL=	LLAMA_BUILD_EXAMPLES
 
 VULKAN_DESC=		Vulkan GPU offload support
 VULKAN_CMAKE_BOOL=	GGML_VULKAN
 VULKAN_BUILD_DEPENDS=	glslc:graphics/shaderc \
 			vulkan-headers>0:graphics/vulkan-headers
 VULKAN_LIB_DEPENDS=	libvulkan.so:graphics/vulkan-loader
 
 BINARY_ALIAS=	git=false
 
 # 2 tests fail: https://github.com/ggerganov/llama.cpp/issues/8906
 
 .include <bsd.port.mk>
diff --git a/misc/llama-cpp/distinfo b/misc/llama-cpp/distinfo
index 1e226726e823..3a7574800002 100644
--- a/misc/llama-cpp/distinfo
+++ b/misc/llama-cpp/distinfo
@@ -1,5 +1,7 @@
-TIMESTAMP = 1724734790
-SHA256 (ggerganov-llama.cpp-b3631_GH0.tar.gz) = 6f2eff255568e757ca7e8a52570be13f60265ec024460b354d0a9cfb3f330573
-SIZE (ggerganov-llama.cpp-b3631_GH0.tar.gz) = 19072652
+TIMESTAMP = 1725086217
+SHA256 (ggerganov-llama.cpp-b3647_GH0.tar.gz) = 03514a396f1366e5f03e181f5f3b9b3a4a595a715b1c505fb6af7674a177a4ed
+SIZE (ggerganov-llama.cpp-b3647_GH0.tar.gz) = 19087310
 SHA256 (nomic-ai-kompute-4565194_GH0.tar.gz) = 95b52d2f0514c5201c7838348a9c3c9e60902ea3c6c9aa862193a212150b2bfc
 SIZE (nomic-ai-kompute-4565194_GH0.tar.gz) = 13540496
+SHA256 (121f915a09c1117d34aff6e8faf6d252aaf11027.patch) = 8e86ce1910aa0eb0512299f95ae2d650a9fbf3818f0c36c7f61799be52db53b1
+SIZE (121f915a09c1117d34aff6e8faf6d252aaf11027.patch) = 589