diff --git a/lib/elixir/lib/kernel/special_forms.ex b/lib/elixir/lib/kernel/special_forms.ex
index e93226a5444..b2856184840 100644
--- a/lib/elixir/lib/kernel/special_forms.ex
+++ b/lib/elixir/lib/kernel/special_forms.ex
@@ -233,9 +233,9 @@ defmodule Kernel.SpecialForms do
 
   Sizes for types are a bit more nuanced. The default size for integers is 8.
 
-  For floats, it is 64. For floats, `size * unit` must result in 32 or 64,
+  For floats, it is 64. For floats, `size * unit` must result in 16, 32, or 64,
   corresponding to [IEEE 754](https://en.wikipedia.org/wiki/IEEE_floating_point)
-  binary32 and binary64, respectively.
+  binary16, binary32, and binary64, respectively.
 
   For binaries, the default is the size of the binary. Only the last binary in a
   match can use the default size. All others must have their size specified
diff --git a/lib/elixir/src/elixir_bitstring.erl b/lib/elixir/src/elixir_bitstring.erl
index 3b653453440..25f88ef307e 100644
--- a/lib/elixir/src/elixir_bitstring.erl
+++ b/lib/elixir/src/elixir_bitstring.erl
@@ -319,8 +319,13 @@ build_spec(Meta, _Size, Unit, Type, _Endianness, Sign, Spec, E) when Type == bin
 build_spec(Meta, Size, Unit, Type, Endianness, Sign, Spec, E) when Type == integer; Type == float ->
   NumberSize = number_size(Size, Unit),
   if
-    Type == float, is_integer(NumberSize), NumberSize /= 32, NumberSize /= 64 ->
-      form_error(Meta, E, ?MODULE, {bittype_float_size, NumberSize});
+    Type == float, is_integer(NumberSize) ->
+      case valid_float_size(NumberSize) of
+        true ->
+          add_spec(Type, add_spec(Endianness, add_spec(Sign, Spec)));
+        false ->
+          form_error(Meta, E, ?MODULE, {bittype_float_size, NumberSize})
+      end;
     Size == default, Unit /= default ->
       form_error(Meta, E, ?MODULE, bittype_unit);
     true ->
@@ -331,6 +336,12 @@ number_size(Size, default) when is_integer(Size) -> Size;
 number_size(Size, Unit) when is_integer(Size) -> Size * Unit;
 number_size(Size, _) -> Size.
 
+%% TODO: Simplify when we require OTP 24
+valid_float_size(16) -> erlang:system_info(otp_release) >= "24";
+valid_float_size(32) -> true;
+valid_float_size(64) -> true;
+valid_float_size(_) -> false.
+
 add_spec(default, Spec) -> Spec;
 add_spec(Key, Spec) -> [{Key, [], []} | Spec].
 
@@ -372,7 +383,12 @@ format_error(bittype_signed) ->
 format_error(bittype_unit) ->
   "integer and float types require a size specifier if the unit specifier is given";
 format_error({bittype_float_size, Other}) ->
-  io_lib:format("float requires size*unit to be 32 or 64 (default), got: ~p", [Other]);
+  Message =
+    case erlang:system_info(otp_release) >= "24" of
+      true -> "16, 32, or 64";
+      false -> "32 or 64"
+    end,
+  io_lib:format("float requires size*unit to be ~s (default), got: ~p", [Message, Other]);
 format_error({invalid_literal, Literal}) ->
   io_lib:format("invalid literal ~ts in <<>>", ['Elixir.Macro':to_string(Literal)]);
 format_error({undefined_bittype, Expr}) ->
diff --git a/lib/elixir/test/elixir/kernel/expansion_test.exs b/lib/elixir/test/elixir/kernel/expansion_test.exs
index 3fbffc1a159..0acba1d2ac7 100644
--- a/lib/elixir/test/elixir/kernel/expansion_test.exs
+++ b/lib/elixir/test/elixir/kernel/expansion_test.exs
@@ -2387,17 +2387,49 @@ defmodule Kernel.ExpansionTest do
       end
     end
 
-    test "raises for invalid size * unit for floats" do
-      message = ~r"float requires size\*unit to be 32 or 64 \(default\), got: 128"
+    # TODO: Simplify when we require OTP 24
+    if System.otp_release() >= "24" do
+      test "16-bit floats" do
+        import Kernel, except: [-: 2]
 
-      assert_raise CompileError, message, fn ->
-        expand(quote(do: <<12.3::32*4>>))
+        assert expand(quote(do: <<12.3::float-16>>)) |> clean_meta([:alignment]) ==
+                 quote(do: <<12.3::float()-size(16)>>)
       end
 
-      message = ~r"float requires size\*unit to be 32 or 64 \(default\), got: 256"
+      test "raises for invalid size * unit for floats" do
+        message = ~r"float requires size\*unit to be 16, 32, or 64 \(default\), got: 128"
 
-      assert_raise CompileError, message, fn ->
-        expand(quote(do: <<12.3::256>>))
+        assert_raise CompileError, message, fn ->
+          expand(quote(do: <<12.3::32*4>>))
+        end
+
+        message = ~r"float requires size\*unit to be 16, 32, or 64 \(default\), got: 256"
+
+        assert_raise CompileError, message, fn ->
+          expand(quote(do: <<12.3::256>>))
+        end
+      end
+    else
+      test "16-bit floats" do
+        message = ~r"float requires size\*unit to be 32 or 64 \(default\), got: 16"
+
+        assert_raise CompileError, message, fn ->
+          expand(quote(do: <<12.3::16>>))
+        end
+      end
+
+      test "raises for invalid size * unit for floats" do
+        message = ~r"float requires size\*unit to be 32 or 64 \(default\), got: 128"
+
+        assert_raise CompileError, message, fn ->
+          expand(quote(do: <<12.3::32*4>>))
+        end
+
+        message = ~r"float requires size\*unit to be 32 or 64 \(default\), got: 256"
+
+        assert_raise CompileError, message, fn ->
+          expand(quote(do: <<12.3::256>>))
+        end
       end
     end