diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8e9e4215..33d151e4 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- feat: update ggml to 0.16.0 by @abetlen in #183
 - feat: add buffer protocol support to utils by @aisk in #175
 
 ## [0.0.45]
diff --git a/ggml/ggml.py b/ggml/ggml.py
index ac0f41ec..3b4ff50d 100644
--- a/ggml/ggml.py
+++ b/ggml/ggml.py
@@ -307,9 +307,9 @@ def f_(*args: Any, **kwargs: Any):
 GGML_EXIT_ABORTED = 1
 
 GGML_VERSION_MAJOR = 0
-GGML_VERSION_MINOR = 15
-GGML_VERSION_PATCH = 3
-GGML_VERSION = "0.15.3"
+GGML_VERSION_MINOR = 16
+GGML_VERSION_PATCH = 0
+GGML_VERSION = "0.16.0"
 
 GGML_ROPE_TYPE_NORMAL = 0
 GGML_ROPE_TYPE_NEOX = 2
@@ -483,7 +483,8 @@ def ggml_fp32_to_bf16_row(fp32: CtypesPointer[ctypes.c_float], bf16: CtypesPoint
 #     GGML_TYPE_MXFP4   = 39,
 #     GGML_TYPE_NVFP4   = 40,
 #     GGML_TYPE_Q1_0    = 41,
-#     GGML_TYPE_COUNT   = 42,
+#     GGML_TYPE_Q2_0    = 42,
+#     GGML_TYPE_COUNT   = 43,
 # };
 GGML_TYPE_F32 = 0
 GGML_TYPE_F16 = 1
@@ -519,7 +520,8 @@ def ggml_fp32_to_bf16_row(fp32: CtypesPointer[ctypes.c_float], bf16: CtypesPoint
 GGML_TYPE_MXFP4 = 39
 GGML_TYPE_NVFP4 = 40
 GGML_TYPE_Q1_0 = 41
-GGML_TYPE_COUNT = 42
+GGML_TYPE_Q2_0 = 42
+GGML_TYPE_COUNT = 43
 
 
 # // precision
@@ -576,6 +578,7 @@ def ggml_fp32_to_bf16_row(fp32: CtypesPointer[ctypes.c_float], bf16: CtypesPoint
 #     GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
 #     GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
 #     GGML_FTYPE_MOSTLY_Q1_0    = 27, // except 1d tensors
+#     GGML_FTYPE_MOSTLY_Q2_0    = 28, // except 1d tensors
 # };
 GGML_FTYPE_UNKNOWN = -1
 GGML_FTYPE_ALL_F32 = 0
@@ -604,6 +607,7 @@ def ggml_fp32_to_bf16_row(fp32: CtypesPointer[ctypes.c_float], bf16: CtypesPoint
 GGML_FTYPE_MOSTLY_MXFP4 = 25
 GGML_FTYPE_MOSTLY_NVFP4 = 26
 GGML_FTYPE_MOSTLY_Q1_0 = 27
+GGML_FTYPE_MOSTLY_Q2_0 = 28
 
 
 # // available tensor operations:
@@ -15783,20 +15787,6 @@ def ggml_backend_cuda_allreduce_tensor(
     ...
 
 
-# GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
-@ggml_function(
-    "ggml_backend_cuda_split_buffer_type",
-    [ctypes.c_int, ctypes.POINTER(ctypes.c_float)],
-    ggml_backend_buffer_type_t_ctypes,
-    enabled=GGML_USE_CUDA,
-)
-def ggml_backend_cuda_split_buffer_type(
-    main_device: Union[ctypes.c_int, int],
-    tensor_split: CtypesArray[ctypes.c_float],
-) -> Optional[ggml_backend_buffer_type_t]:
-    ...
-
-
 # // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 # GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 @ggml_function(
diff --git a/vendor/ggml b/vendor/ggml
index eced84c8..524f974b 160000
--- a/vendor/ggml
+++ b/vendor/ggml
@@ -1 +1 @@
-Subproject commit eced84c86f8b012c752c016f7fe789adea168e1e
+Subproject commit 524f974bb21a1013408f76d71c15732482c0c3fe