diff --git a/backend/cpp/llama-cpp/Makefile b/backend/cpp/llama-cpp/Makefile
index 8b145e4519f1..5f285ab217c8 100644
--- a/backend/cpp/llama-cpp/Makefile
+++ b/backend/cpp/llama-cpp/Makefile
@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=ced765be44ce173c374f295b3c6f4175f8fd109b
+LLAMA_VERSION?=706e3f93a60109a40f1224eaf4af0d59caa7c3ae
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=
diff --git a/backend/cpp/llama-cpp/grpc-server.cpp b/backend/cpp/llama-cpp/grpc-server.cpp
index 3f33c74bf271..1009d36fd7df 100644
--- a/backend/cpp/llama-cpp/grpc-server.cpp
+++ b/backend/cpp/llama-cpp/grpc-server.cpp
@@ -358,9 +358,7 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
 
     params.model.path = request->modelfile();
     if (!request->mmproj().empty()) {
-    // get the directory of modelfile
-      std::string model_dir = params.model.path.substr(0, params.model.path.find_last_of("/\\"));
-      params.mmproj.path = model_dir + "/"+ request->mmproj();
+      params.mmproj.path = request->mmproj();
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
diff --git a/core/backend/options.go b/core/backend/options.go
index b585a22b3cd2..f3d5a4ccd402 100644
--- a/core/backend/options.go
+++ b/core/backend/options.go
@@ -36,7 +36,7 @@ func ModelOptions(c config.ModelConfig, so *config.ApplicationConfig, opts ...mo
 
 	c.Threads = &threads
 
-	grpcOpts := grpcModelOpts(c)
+	grpcOpts := grpcModelOpts(c, so.SystemState.Model.ModelsPath)
 	defOpts = append(defOpts, model.WithLoadGRPCLoadModelOpts(grpcOpts))
 
 	if so.ParallelBackendRequests {
@@ -72,7 +72,7 @@ func getSeed(c config.ModelConfig) int32 {
 	return seed
 }
 
-func grpcModelOpts(c config.ModelConfig) *pb.ModelOptions {
+func grpcModelOpts(c config.ModelConfig, modelPath string) *pb.ModelOptions {
 	b := 512
 	if c.Batch != 0 {
 		b = c.Batch
@@ -131,7 +131,7 @@ func grpcModelOpts(c config.ModelConfig) *pb.ModelOptions {
 		})
 	}
 
-	return &pb.ModelOptions{
+	opts := &pb.ModelOptions{
 		CUDA:                 c.CUDA || c.Diffusers.CUDA,
 		SchedulerType:        c.Diffusers.SchedulerType,
 		GrammarTriggers:      triggers,
@@ -170,7 +170,6 @@ func grpcModelOpts(c config.ModelConfig) *pb.ModelOptions {
 		LimitImagePerPrompt: int32(c.LimitMMPerPrompt.LimitImagePerPrompt),
 		LimitVideoPerPrompt: int32(c.LimitMMPerPrompt.LimitVideoPerPrompt),
 		LimitAudioPerPrompt: int32(c.LimitMMPerPrompt.LimitAudioPerPrompt),
-		MMProj:              c.MMProj,
 		FlashAttention:      flashAttention,
 		CacheTypeKey:        c.CacheTypeK,
 		CacheTypeValue:      c.CacheTypeV,
@@ -198,6 +197,12 @@ func grpcModelOpts(c config.ModelConfig) *pb.ModelOptions {
 		// RWKV
 		Tokenizer: c.Tokenizer,
 	}
+
+	if c.MMProj != "" {
+		opts.MMProj = filepath.Join(modelPath, c.MMProj)
+	}
+
+	return opts
 }
 
 func gRPCPredictOpts(c config.ModelConfig, modelPath string) *pb.PredictOptions {