diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
index 0f2b4d5..811f69a 100644
--- a/llama_cpp/llama_cpp.py
+++ b/llama_cpp/llama_cpp.py
@@ -114,6 +114,7 @@ LLAMA_FTYPE_ALL_F32 = ctypes.c_int(0)
 LLAMA_FTYPE_MOSTLY_F16 = ctypes.c_int(1)  # except 1d tensors
 LLAMA_FTYPE_MOSTLY_Q4_0 = ctypes.c_int(2)  # except 1d tensors
 LLAMA_FTYPE_MOSTLY_Q4_1 = ctypes.c_int(3)  # except 1d tensors
+LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = ctypes.c_int(4)  # tok_embeddings.weight and output.weight are F16
 
 # Functions
 
diff --git a/vendor/llama.cpp b/vendor/llama.cpp
index 8b67998..e7f6997 160000
--- a/vendor/llama.cpp
+++ b/vendor/llama.cpp
@@ -1 +1 @@
-Subproject commit 8b679987cdce292ff36bd741f6715e4927e26f9b
+Subproject commit e7f6997f897a18b6372a6460e25c5f89e1469f1d