diff --git a/exla/lib/exla/defn.ex b/exla/lib/exla/defn.ex
index 413c38ce45..7a7b3865a8 100644
--- a/exla/lib/exla/defn.ex
+++ b/exla/lib/exla/defn.ex
@@ -1209,6 +1209,10 @@ defmodule EXLA.Defn do
     EXLA.Lib.argsort(state.builder, tensor, dimension, stable, comp, ans.type)
   end
 
+  defp to_operator(:elixir_call, _, _, _) do
+    raise "Nx.elixir_call/3 is not supported yet. Use Nx.Defn.Evaluator as your compiler."
+  end
+
   defp fft(exla_op, [%Value{} = tensor, opts], %{type: type} = ans, state) do
     n = opts[:length]
     axis = opts[:axis]
diff --git a/exla/test/exla/defn/elixir_call_test.exs b/exla/test/exla/defn/elixir_call_test.exs
new file mode 100644
index 0000000000..add051a3f6
--- /dev/null
+++ b/exla/test/exla/defn/elixir_call_test.exs
@@ -0,0 +1,61 @@
+defmodule EXLA.Defn.ElixirCallEvaluatorTest do
+  use ExUnit.Case, async: true
+  import Nx.Defn
+  import Nx.Testing
+
+  setup do
+    Nx.Defn.default_options(compiler: Nx.Defn.Evaluator)
+    Nx.default_backend(EXLA.Backend)
+    :ok
+  end
+
+  defn add_offset(x) do
+    out = %{x | type: Nx.Type.to_floating(x.type)}
+
+    Nx.elixir_call(out, [x, [offset: 10.0]], fn t, opts ->
+      Nx.add(Nx.as_type(t, :f32), opts[:offset])
+    end)
+  end
+
+  test "elixir_call with single output" do
+    x = Nx.iota({5})
+    y = add_offset(x)
+
+    expected = Nx.add(Nx.as_type(x, :f32), 10.0)
+    assert_equal(y, expected)
+  end
+
+  defn split_and_sum(x) do
+    fx = Nx.as_type(x, :f32)
+
+    out0 = fx
+    out1 = fx
+    out_template = {out0, out1}
+
+    {a, b} =
+      Nx.elixir_call(out_template, [fx], fn t ->
+        {Nx.multiply(t, 2.0), Nx.add(t, 1.0)}
+      end)
+
+    Nx.add(a, b)
+  end
+
+  test "elixir_call with tuple output" do
+    x = Nx.tensor([1, 2, 3])
+    y = split_and_sum(x)
+
+    fx = Nx.as_type(x, :f32)
+    expected = Nx.add(Nx.multiply(fx, 2.0), Nx.add(fx, 1.0))
+    assert_equal(y, expected)
+  end
+
+  test "fails when using EXLA compiler" do
+    x = Nx.tensor([1, 2, 3])
+
+    assert_raise RuntimeError,
+                 "Nx.elixir_call/3 is not supported yet. Use Nx.Defn.Evaluator as your compiler.",
+                 fn ->
+                   EXLA.jit_apply(&split_and_sum/1, [x])
+                 end
+  end
+end
diff --git a/nx/lib/nx.ex b/nx/lib/nx.ex
index 091372d005..715a149286 100644
--- a/nx/lib/nx.ex
+++ b/nx/lib/nx.ex
@@ -2196,6 +2196,61 @@ defmodule Nx do
     list
   end
 
+  @doc """
+  Invokes an Elixir function from within defn.
+
+  This function allows integrating arbitrary Elixir code into `defn` graphs.
+  It receives an output template (a tensor or a tuple of tensors) that
+  specifies the expected shapes, types, and names of the result, a list of
+  arguments to pass to the Elixir function, and the function itself.
+
+  Inside `defn`, this builds an expression node understood by compilers.
+  Outside `defn` or on backends without special support, it executes `fun`
+  directly and validates the result matches the template.
+  """
+  @doc type: :backend
+  def elixir_call(output, args, fun) when is_list(args) and is_function(fun) do
+    {:arity, arity} = Function.info(fun, :arity)
+    num_args = length(args)
+
+    if arity != num_args do
+      raise ArgumentError,
+            "expected #{arity} arguments, got #{num_args}"
+    end
+
+    backend = Nx.Shared.list_impl!(args)
+
+    cond do
+      function_exported?(backend, :elixir_call, 3) ->
+        output
+        |> backend.elixir_call(args, fun)
+        |> ensure_call_compatible!(output)
+
+      true ->
+        fun
+        |> apply(args)
+        |> ensure_call_compatible!(output)
+    end
+  end
+
+  defp ensure_call_compatible!(left, right) when tuple_size(left) == tuple_size(right) do
+    [Tuple.to_list(left), Tuple.to_list(right)]
+    |> Enum.zip_with(fn [l, r] -> ensure_call_compatible!(l, r) end)
+
+    left
+  end
+
+  defp ensure_call_compatible!(
+         %{shape: shape, type: type, names: names} = left,
+         %{shape: shape, type: type, names: names}
+       ),
+       do: left
+
+  defp ensure_call_compatible!(left, right) do
+    raise ArgumentError,
+          "expected the elixir_call function to match the given output template #{inspect(right)}, got: #{inspect(left)}"
+  end
+
   defp chunk([], data, type) do
     match_types [type] do
       <<match!(head, 0), tail::binary>> = data
diff --git a/nx/lib/nx/backend.ex b/nx/lib/nx/backend.ex
index 3c463ba237..f8556ce308 100644
--- a/nx/lib/nx/backend.ex
+++ b/nx/lib/nx/backend.ex
@@ -142,6 +142,15 @@ defmodule Nx.Backend do
   """
   @callback optional(atom, [term], fun) :: tensor
 
+  @doc """
+  Invoked to execute a generic Elixir callback from within defn.
+
+  The backend may choose how to execute it. For example, EXLA can lower
+  to a custom_call that interacts with Erlang/Elixir via C; pure CPU
+  backends may call the function directly.
+  """
+  @callback elixir_call(out :: tensor | tuple, [term], fun) :: tensor
+
   @callback qr({q :: tensor, r :: tensor}, tensor, keyword) :: tensor
   @callback cholesky(out :: tensor, tensor) :: tensor
   @callback eigh({eigenvals :: tensor, eigenvecs :: tensor}, tensor, keyword) :: tensor
@@ -162,6 +171,7 @@ defmodule Nx.Backend do
 
   @optional_callbacks [
     optional: 3,
+    elixir_call: 3,
     solve: 3,
     determinant: 2,
     logical_not: 2,
diff --git a/nx/lib/nx/defn/evaluator.ex b/nx/lib/nx/defn/evaluator.ex
index d028ec6a63..c913f4ec3c 100644
--- a/nx/lib/nx/defn/evaluator.ex
+++ b/nx/lib/nx/defn/evaluator.ex
@@ -175,6 +175,15 @@ defmodule Nx.Defn.Evaluator do
     Map.put(cache, [:optional | id], optional_expr_cache)
   end
 
+  defp compute_cache(:elixir_call, %{data: %Expr{args: args}}, state, cache) do
+    [in_args, _fun] = args
+
+    Enum.reduce(in_args, cache, fn
+      t, cache when is_list(t) -> cache
+      t, cache -> compute_cache(t, state, cache)
+    end)
+  end
+
   defp compute_cache(:cond, %{data: %Expr{args: [clauses, last], id: id}}, state, cache) do
     %{parent_ids: parent_ids, current_ids: current_ids} = state
 
@@ -431,6 +440,23 @@ defmodule Nx.Defn.Evaluator do
     end
   end
 
+  defp eval_apply(
+         :elixir_call,
+         %{data: %Expr{args: [in_args, fun]}} = expr,
+         state,
+         caches
+       ) do
+    {tensor_args, opts} = Enum.split_while(in_args, &(not is_list(&1)))
+    {evaluated_tensors, caches} = Enum.map_reduce(tensor_args, caches, &eval(&1, state, &2))
+    backend = Nx.Shared.list_impl!(evaluated_tensors)
+
+    if backend == Nx.Defn.Expr do
+      {expr, caches}
+    else
+      {apply(fun, evaluated_tensors ++ opts), caches}
+    end
+  end
+
   defp eval_apply(op, %{vectorized_axes: [_ | _]} = ans, _state, _caches) do
     raise "unexpected vectorized axes in evaluator for operation #{inspect(op)}: #{inspect(ans)}"
   end
diff --git a/nx/lib/nx/defn/expr.ex b/nx/lib/nx/defn/expr.ex
index 782e4a07fd..1d488df888 100644
--- a/nx/lib/nx/defn/expr.ex
+++ b/nx/lib/nx/defn/expr.ex
@@ -41,6 +41,8 @@ defmodule Nx.Defn.Expr do
 
     * `attach_token(token(%Nx.Defn.Token{}), expr)`
 
+    * `elixir_call(name, args, fun)`
+
   `defn` compilers must handle said nodes accordingly.
   """
 
@@ -384,6 +386,22 @@ defmodule Nx.Defn.Expr do
     end
   end
 
+  @impl true
+  def elixir_call(out, in_args, fun) do
+    {tensor_args, _opts} = Enum.split_while(in_args, &(not is_list(&1)))
+    [%T{data: %Expr{context: context}} | _] = Enum.map(tensor_args, &to_expr/1)
+
+    case out do
+      t when is_struct(t, Nx.Tensor) ->
+        expr(t, context, :elixir_call, [in_args, fun])
+
+      tuple when is_tuple(tuple) ->
+        out_template = tuple_out(tuple_size(tuple))
+        expr_node = expr(out_template, context, :elixir_call, [in_args, fun])
+        tuple(expr_node, Tuple.to_list(tuple))
+    end
+  end
+
   ## Nx.Defn AST callbacks
 
   @doc false
diff --git a/nx/lib/nx/defn/grad.ex b/nx/lib/nx/defn/grad.ex
index 2941889f98..8c72d0fed0 100644
--- a/nx/lib/nx/defn/grad.ex
+++ b/nx/lib/nx/defn/grad.ex
@@ -122,6 +122,10 @@ defmodule Nx.Defn.Grad do
     acc
   end
 
+  defp parents_args(:elixir_call, _expr, _id, acc, _parent_vectorized_names) do
+    acc
+  end
+
   defp parents_args(
          :optional,
          %{data: %{args: [call, _expr, callback]}} = t,
diff --git a/nx/lib/nx/defn/tree.ex b/nx/lib/nx/defn/tree.ex
index 582b9d4689..733a131e4f 100644
--- a/nx/lib/nx/defn/tree.ex
+++ b/nx/lib/nx/defn/tree.ex
@@ -192,6 +192,21 @@ defmodule Nx.Defn.Tree do
     {[call, expr, callback], acc}
   end
 
+  def apply_args(%T{data: %Expr{op: :elixir_call, args: args}}, _type, acc, fun) do
+    [in_args, callback] = args
+
+    {in_args, acc} =
+      Enum.map_reduce(in_args, acc, fn t, acc ->
+        if is_list(t) do
+          {t, acc}
+        else
+          Composite.traverse(t, acc, fun)
+        end
+      end)
+
+    {[in_args, callback], acc}
+  end
+
   def apply_args(%T{data: %Expr{op: :token, args: [token]}}, _type, acc, fun) do
     {hooks, acc} =
       Enum.map_reduce(token.hooks, acc, fn %{expr: expr} = token, acc ->
diff --git a/nx/test/nx/defn/elixir_call_evaluator_test.exs b/nx/test/nx/defn/elixir_call_evaluator_test.exs
new file mode 100644
index 0000000000..92fad6b431
--- /dev/null
+++ b/nx/test/nx/defn/elixir_call_evaluator_test.exs
@@ -0,0 +1,49 @@
+defmodule Nx.Defn.ElixirCallEvaluatorTest do
+  use ExUnit.Case, async: true
+  import Nx.Defn
+
+  setup do
+    Nx.Defn.default_options(compiler: Nx.Defn.Evaluator)
+    :ok
+  end
+
+  defn add_offset(x) do
+    out = %{x | type: Nx.Type.to_floating(x.type)}
+
+    Nx.elixir_call(out, [x, [offset: 10.0]], fn t, opts ->
+      Nx.add(Nx.as_type(t, :f32), opts[:offset])
+    end)
+  end
+
+  test "elixir_call with single output" do
+    x = Nx.iota({5})
+    y = add_offset(x)
+
+    expected = Nx.add(Nx.as_type(x, :f32), 10.0)
+    assert Nx.all_close(y, expected) |> Nx.to_number() == 1
+  end
+
+  defn split_and_sum(x) do
+    fx = Nx.as_type(x, :f32)
+
+    out0 = fx
+    out1 = fx
+    out_template = {out0, out1}
+
+    {a, b} =
+      Nx.elixir_call(out_template, [fx], fn t ->
+        {Nx.multiply(t, 2.0), Nx.add(t, 1.0)}
+      end)
+
+    Nx.add(a, b)
+  end
+
+  test "elixir_call with tuple output" do
+    x = Nx.tensor([1, 2, 3])
+    y = split_and_sum(x)
+
+    fx = Nx.as_type(x, :f32)
+    expected = Nx.add(Nx.multiply(fx, 2.0), Nx.add(fx, 1.0))
+    assert expected == y
+  end
+end
diff --git a/torchx/mix.exs b/torchx/mix.exs
index fa5531e541..5174e59cdb 100644
--- a/torchx/mix.exs
+++ b/torchx/mix.exs
@@ -41,8 +41,8 @@ defmodule Torchx.MixProject do
 
   defp deps do
     [
-      {:nx, "~> 0.10.0"},
-      # {:nx, path: "../nx"},
+      # {:nx, "~> 0.10.0"},
+      {:nx, path: "../nx"},
       {:ex_doc, "~> 0.29", only: :docs}
     ]
   end
diff --git a/torchx/test/torchx/defn/elixir_call_test.exs b/torchx/test/torchx/defn/elixir_call_test.exs
new file mode 100644
index 0000000000..9c504fa6c8
--- /dev/null
+++ b/torchx/test/torchx/defn/elixir_call_test.exs
@@ -0,0 +1,51 @@
+defmodule Torchx.Defn.ElixirCallEvaluatorTest do
+  use ExUnit.Case, async: true
+  import Nx.Defn
+  import Nx.Testing
+
+  setup do
+    Nx.Defn.default_options(compiler: Nx.Defn.Evaluator)
+    Nx.default_backend(Torchx.Backend)
+    :ok
+  end
+
+  defn add_offset(x) do
+    out = %{x | type: Nx.Type.to_floating(x.type)}
+
+    Nx.elixir_call(out, [x, [offset: 10.0]], fn t, opts ->
+      Nx.add(Nx.as_type(t, :f32), opts[:offset])
+    end)
+  end
+
+  test "elixir_call with single output" do
+    x = Nx.iota({5})
+    y = add_offset(x)
+
+    expected = Nx.add(Nx.as_type(x, :f32), 10.0)
+    assert_equal(y, expected)
+  end
+
+  defn split_and_sum(x) do
+    fx = Nx.as_type(x, :f32)
+
+    out0 = fx
+    out1 = fx
+    out_template = {out0, out1}
+
+    {a, b} =
+      Nx.elixir_call(out_template, [fx], fn t ->
+        {Nx.multiply(t, 2.0), Nx.add(t, 1.0)}
+      end)
+
+    Nx.add(a, b)
+  end
+
+  test "elixir_call with tuple output" do
+    x = Nx.tensor([1, 2, 3])
+    y = split_and_sum(x)
+
+    fx = Nx.as_type(x, :f32)
+    expected = Nx.add(Nx.multiply(fx, 2.0), Nx.add(fx, 1.0))
+    assert_equal(y, expected)
+  end
+end