diff --git a/src/panfrost/compiler/kraid/compile.rs b/src/panfrost/compiler/kraid/compile.rs
index 53d9d23992f..b5e3ceaaee1 100644
--- a/src/panfrost/compiler/kraid/compile.rs
+++ b/src/panfrost/compiler/kraid/compile.rs
@@ -81,6 +81,10 @@ pub extern "C" fn kraid_compile_nir(
     dump_shader(&s, "after register assignment");
     s.validate();
 
+    s.lower_16bit_alu();
+    dump_shader(&s, "after lowering 16bit ALU ops");
+    s.validate();
+
     s.assign_message_slots();
     dump_shader(&s, "after message slot assignment");
     s.validate();
diff --git a/src/panfrost/compiler/kraid/lib.rs b/src/panfrost/compiler/kraid/lib.rs
index b6cd1f694ff..9ab415a14da 100644
--- a/src/panfrost/compiler/kraid/lib.rs
+++ b/src/panfrost/compiler/kraid/lib.rs
@@ -9,6 +9,7 @@ mod encode_v9;
 mod flow;
 mod ir;
 mod isa;
+mod lower_16bit;
 mod message_slots;
 mod model;
 mod nir;
diff --git a/src/panfrost/compiler/kraid/lower_16bit.rs b/src/panfrost/compiler/kraid/lower_16bit.rs
new file mode 100644
index 00000000000..9e0ef1e58b1
--- /dev/null
+++ b/src/panfrost/compiler/kraid/lower_16bit.rs
@@ -0,0 +1,76 @@
+// Copyright © 2026 Collabora, Ltd.
+// SPDX-License-Identifier: MIT
+
+use crate::ir::*;
+use crate::ops::*;
+use std::num::NonZeroU8;
+
+fn type_is_16bit(data_type: DataType) -> bool {
+    data_type.total_bits() == std::num::NonZeroU8::new(16)
+}
+
+fn replicate_type(data_type: DataType, n: u8) -> DataType {
+    DataType::v(
+        data_type.comps().unwrap().get() * n,
+        data_type.scalar_type(),
+    )
+}
+
+macro_rules! lower_op {
+    ($op: expr, $variant: ident) => {{
+        const SOME_8: Option<NonZeroU8> = NonZeroU8::new(8);
+        const SOME_16: Option<NonZeroU8> = NonZeroU8::new(16);
+        const SOME_32: Option<NonZeroU8> = NonZeroU8::new(32);
+        match $op.$variant.total_bits() {
+            SOME_8 => {
+                for src in $op.srcs() {
+                    debug_assert!(src.swizzle.replicates_byte());
+                }
+                $op.$variant = replicate_type($op.$variant, 4);
+            }
+            SOME_16 => {
+                for src in $op.srcs() {
+                    debug_assert!(src.swizzle.replicates_half());
+                }
+                $op.$variant = replicate_type($op.$variant, 2);
+            }
+            bits => assert!(bits >= SOME_32),
+        }
+    }};
+}
+
+fn lower_instr(instr: &mut Instr) {
+    match &mut instr.op {
+        Op::FAdd(op) => lower_op!(op, dst_type),
+        Op::FCmp(op) => lower_op!(op, src_type),
+        Op::IAdd(op) => lower_op!(op, dst_type),
+        Op::ICmp(op) => lower_op!(op, src_type),
+        Op::LdPka(_) | Op::Load(_) => (), // These handle 16-bit natively
+        Op::Mov(op) => {
+            if op.dst_type.total_bits() == std::num::NonZeroU8::new(16) {
+                instr.op = Op::from(OpIAdd {
+                    dst: op.dst.clone(),
+                    dst_type: DataType::V2I16,
+                    saturate: false,
+                    srcs: [0.into(), op.src.clone().half(0)],
+                });
+            }
+        }
+        Op::ShiftLop(op) => lower_op!(op, dst_type),
+        op => {
+            for dst in op.dsts() {
+                assert!(dst.bytes_written() >= 4);
+            }
+        }
+    }
+}
+
+impl Shader<'_> {
+    pub fn lower_16bit_alu(&mut self) {
+        for b in self.blocks.iter_mut() {
+            for i in b.instrs.iter_mut() {
+                lower_instr(i);
+            }
+        }
+    }
+}