diff --git a/.github/workflows/emlx.yml b/.github/workflows/emlx.yml
index 6e50dad..8fd1392 100644
--- a/.github/workflows/emlx.yml
+++ b/.github/workflows/emlx.yml
@@ -72,7 +72,7 @@ jobs:
           mix test --warnings-as-errors
 
   macos:
-    name: macOS (${{ matrix.job.elixir }}, ${{ matrix.job.otp }})
+    name: macOS ${{ matrix.job.gpu && 'gpu' || 'cpu' }} (${{ matrix.job.elixir }}, ${{ matrix.job.otp }})
     runs-on: macos-26
     strategy:
       fail-fast: false
diff --git a/emlx/c_src/emlx_nif.cpp b/emlx/c_src/emlx_nif.cpp
index b62948d..f18a408 100644
--- a/emlx/c_src/emlx_nif.cpp
+++ b/emlx/c_src/emlx_nif.cpp
@@ -218,7 +218,7 @@ NIF(to_blob) {
 }
 
 uint64_t elem_count(std::vector<int> shape) {
-  return std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<>{});
+  return std::accumulate(shape.begin(), shape.end(), uint64_t{1}, std::multiplies<uint64_t>{});
 }
 
 NIF(from_blob) {
diff --git a/emlx/test/emlx_test.exs b/emlx/test/emlx_test.exs
index 53cc1ae..8d2a33a 100644
--- a/emlx/test/emlx_test.exs
+++ b/emlx/test/emlx_test.exs
@@ -170,4 +170,21 @@ defmodule EMLXTest do
       end
     end
   end
+
+  # Regression: elem_count overflowed int32 for shapes whose element count
+  # exceeds INT32_MAX, causing "Binary size is too small" on valid binaries.
+  @tag :large_memory
+  test "from_binary accepts shape whose element count exceeds INT32_MAX" do
+    # Reshape on BinaryBackend first — Nx.from_binary creates a flat 1D tensor
+    # whose single dimension would also exceed INT32_MAX.
+    int_32_max = 2 ** 31
+    binary = String.duplicate(<<7>>, int_32_max)
+    shape = {2, div(int_32_max, 2)}
+
+    out = Nx.template(shape, :u8)
+    t = EMLX.Backend.from_binary(out, binary, backend: EMLX.Backend)
+
+    assert Nx.shape(t) == shape
+    assert_equal(Nx.all(Nx.equal(t, 7)), 1)
+  end
 end
diff --git a/emlx/test/test_helper.exs b/emlx/test/test_helper.exs
index 4cf03a2..a6ed859 100644
--- a/emlx/test/test_helper.exs
+++ b/emlx/test/test_helper.exs
@@ -51,7 +51,7 @@ distributed_exclude =
 
 gpu_exclude =
   case EMLX.NIF.command_queue_new(:gpu) do
-    {:ok, _} -> []
+    {:ok, _} -> [:large_memory]
     {:error, _} -> [:metal]
   end