diff --git a/Dockerfile b/Dockerfile
index 566e03bc6265..2f2bcafa9f06 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -303,7 +303,7 @@ RUN make prepare
 ## We only leave the most CPU-optimized variant and the fallback for the cublas/hipblas build
 ## (both will use CUDA or hipblas for the actual computation)
 RUN if [ "${BUILD_TYPE}" = "cublas" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
-        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
+        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx512 backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
     else \
         make build; \
     fi
diff --git a/Makefile b/Makefile
index 9c4f3778225a..e3c280396943 100644
--- a/Makefile
+++ b/Makefile
@@ -186,6 +186,7 @@ endif
 ALL_GRPC_BACKENDS=backend-assets/grpc/huggingface
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx2
+ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx512
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-fallback
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-ggml
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-grpc
@@ -699,6 +700,13 @@ backend-assets/grpc/llama-cpp-avx2: backend-assets/grpc backend/cpp/llama/llama.
 	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on" $(MAKE) VARIANT="llama-avx2" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-avx2/grpc-server backend-assets/grpc/llama-cpp-avx2
 
+backend-assets/grpc/llama-cpp-avx512: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-avx512
+	$(MAKE) -C backend/cpp/llama-avx512 purge
+	$(info ${GREEN}I llama-cpp build info:avx512${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on" $(MAKE) VARIANT="llama-avx512" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-avx512/grpc-server backend-assets/grpc/llama-cpp-avx512
+
 backend-assets/grpc/llama-cpp-avx: backend-assets/grpc backend/cpp/llama/llama.cpp
 	cp -rf backend/cpp/llama backend/cpp/llama-avx
 	$(MAKE) -C backend/cpp/llama-avx purge
diff --git a/pkg/model/initializers.go b/pkg/model/initializers.go
index 9fc0c18c089d..ace72fa3d342 100644
--- a/pkg/model/initializers.go
+++ b/pkg/model/initializers.go
@@ -48,6 +48,7 @@ const (
 	LLamaCPP = "llama-cpp"
 
 	LLamaCPPAVX2     = "llama-cpp-avx2"
+	LLamaCPPAVX512   = "llama-cpp-avx512"
 	LLamaCPPAVX      = "llama-cpp-avx"
 	LLamaCPPFallback = "llama-cpp-fallback"
 	LLamaCPPCUDA     = "llama-cpp-cuda"
@@ -68,6 +69,7 @@ const (
 
 var llamaCPPVariants = []string{
 	LLamaCPPAVX2,
+	LLamaCPPAVX512,
 	LLamaCPPAVX,
 	LLamaCPPFallback,
 	LLamaCPPCUDA,
@@ -268,6 +270,12 @@ func selectGRPCProcessByHostCapabilities(backend, assetDir string, f16 bool) str
 			log.Info().Msgf("[%s] attempting to load with AVX2 variant", backend)
 			selectedProcess = p
 		}
+	} else if xsysinfo.HasCPUCaps(cpuid.AVX512F) {
+		p := backendPath(assetDir, LLamaCPPAVX512)
+		if _, err := os.Stat(p); err == nil {
+			log.Info().Msgf("[%s] attempting to load with AVX512 variant", backend)
+			selectedProcess = p
+		}
 	} else if xsysinfo.HasCPUCaps(cpuid.AVX) {
 		p := backendPath(assetDir, LLamaCPPAVX)
 		if _, err := os.Stat(p); err == nil {