diff --git a/include/tvm/topi/transform.h b/include/tvm/topi/transform.h
index 7092eaa42ccb..f41cc7a12e91 100644
--- a/include/tvm/topi/transform.h
+++ b/include/tvm/topi/transform.h
@@ -1606,11 +1606,16 @@ inline Tensor gather_nd(const Tensor& data, const Tensor& indices, int batch_dim
           // Index tensors are validated by integer element kind; vector lane encoding is
           // irrelevant for choosing whether an index cast is needed.
           PrimType indices_ty = indices->dtype;
-          if (indices_ty.MatchesCode(DLDataTypeCode::kDLInt, DLDataTypeCode::kDLUInt)) {
-            real_indices.push_back(indices(indices_position));
-          } else {
-            real_indices.push_back(tvm::cast(tvm::PrimType::Int(32), indices(indices_position)));
+          PrimExpr idx = indices(indices_position);
+          if (!indices_ty.MatchesCode(DLDataTypeCode::kDLInt, DLDataTypeCode::kDLUInt)) {
+            idx = tvm::cast(tvm::PrimType::Int(32), idx);
           }
+          // Support negative indices like the ONNX GatherND op does (and
+          // topi.scatter_elements already does): an in-range negative index
+          // counts from the end of the corresponding data axis.
+          PrimExpr axis_size = data->shape[batch_dims + i];
+          idx = idx + tvm::cast(tvm::PrimType::Int(32), idx < 0) * axis_size;
+          real_indices.push_back(idx);
         }
         if (real_indices.size() == ndim_d) {
           return data(real_indices);
diff --git a/tests/python/relax/test_op_manipulate.py b/tests/python/relax/test_op_manipulate.py
index 694ec82baf90..429bf7f597d6 100644
--- a/tests/python/relax/test_op_manipulate.py
+++ b/tests/python/relax/test_op_manipulate.py
@@ -17,9 +17,10 @@
 # ruff: noqa: E731, F841
 import pytest
 
+import numpy as np
 import tvm
 import tvm.testing
-from tvm import relax, tirx
+from tvm import relax, te, tirx, topi
 from tvm.ir import Op, VDevice
 from tvm.script import relax as R
 from tvm.script import tirx as T
@@ -3209,6 +3210,33 @@ def test_gather_nd_infer_ty_wrong_inputs():
         bb.normalize(relax.op.gather_nd(x0, i1))
 
 
+@pytest.mark.parametrize("batch_dims", [0, 1])
+def test_gather_nd_negative_indices_topi(batch_dims):
+    # topi.gather_nd supports in-range negative indices (ONNX GatherND
+    # semantics) even when the indices are runtime values, matching what
+    # topi.scatter_elements already does.
+    data_np = np.arange(24, dtype="int32").reshape(2, 3, 4)
+    if batch_dims == 0:
+        # index row k addresses data axis k (sizes 2, 3, 4)
+        indices_np = np.array(
+            [[[0, -1], [-1, 1]], [[-2, 1], [2, -2]], [[-4, 3], [3, -1]]], dtype="int64"
+        )
+        expected = data_np[tuple(indices_np)]
+    else:
+        indices_np = np.array([[[-1, -2], [-3, 1]], [[2, -1], [0, -3]]], dtype="int64")
+        expected = np.stack([data_np[b][tuple(indices_np[:, b])] for b in range(2)])
+
+    data = te.placeholder(tuple(data_np.shape), dtype="int32", name="data")
+    indices = te.placeholder(tuple(indices_np.shape), dtype="int64", name="indices")
+    out = topi.gather_nd(data, indices, batch_dims)
+    func = te.create_prim_func([data, indices, out]).with_attr("target", tvm.target.Target("llvm"))
+    f = tvm.build(tvm.IRModule({"main": func}), target="llvm")
+
+    out_t = tvm.runtime.tensor(np.zeros(expected.shape, dtype="int32"))
+    f(tvm.runtime.tensor(data_np), tvm.runtime.tensor(indices_np), out_t)
+    tvm.testing.assert_allclose(out_t.numpy(), expected)
+
+
 def test_scatter_elements_infer_ty():
     bb = relax.BlockBuilder()
     vdev0 = VDevice("llvm")