diff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cpp
index e873024c90..62f83e3ea5 100644
--- a/src/models/kimi-linear.cpp
+++ b/src/models/kimi-linear.cpp
@@ -3,7 +3,8 @@
 
 #define CHUNK_SIZE 64
 
-llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params), model(model) {
+llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context_mamba(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
diff --git a/src/models/models.h b/src/models/models.h
index 3ed00aae32..549329e15a 100644
--- a/src/models/models.h
+++ b/src/models/models.h
@@ -287,7 +287,6 @@ struct llm_build_jamba : public llm_graph_context_mamba {
 struct llm_build_kimi_linear : public llm_graph_context_mamba {
     llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params);
 private:
-    const llama_model & model;
     ggml_tensor * build_kda_autoregressive(
                 ggml_tensor * q,
                 ggml_tensor * k,
@@ -308,6 +307,8 @@ private:
                 ggml_tensor * identity,
                 ggml_tensor * diag_mask,
                         int   il);
+
+    const llama_model & model;
 };
 
 struct llm_build_lfm2 : public llm_graph_context {