diff --git a/common/arg.cpp b/common/arg.cpp
index fcee0c4470077..7507c81155071 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2466,7 +2466,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
-        "number of layers to store in VRAM",
+        string_format("max. number of layers to store in VRAM (default: %d)", params.n_gpu_layers),
         [](common_params & params, int value) {
             params.n_gpu_layers = value;
             if (!llama_supports_gpu_offload()) {