diff --git a/docs/tutorial.md b/docs/tutorial.md
index 4a356865..a9312c1f 100644
--- a/docs/tutorial.md
+++ b/docs/tutorial.md
@@ -1365,6 +1365,8 @@ This example brings together structs, `f32` math, tuples, and functions — all
 
 **Modulo (`%`)** gives you the remainder left over after a division. For example, `10 % 3` is `1`, because 10 divided by 3 is 3 with 1 left over. In DSP this is useful for things like wrapping a value around a range — keeping a phase counter cycling between 0 and some maximum, for instance.
 
+`%` works on floats too, and there it behaves like C's `fmod`: the result takes the sign of the left operand, so `-0.25 % 1.0` is `-0.25`. If you want the result to take the sign of the right operand instead (so that `mod(-0.25, 1.0)` is `0.75`), use the stdlib's `mod(x, y)`.
+
 **Power (`^`)** raises a number to an exponent. For example, `2 ^ 8` is 256 (2 multiplied by itself 8 times).
 
 ### Operator precedence
@@ -1536,7 +1538,7 @@ Here are some common math helpers from `stdlib.lyte`:
 | Function | What it does |
 |----------|--------------|
 | `fract(x)` | Fractional part of `x` |
-| `mod(x, y)` | Floating-point modulo |
+| `mod(x, y)` | Floating-point modulo, result takes the sign of `y` (`%` takes the sign of `x`) |
 | `clamp(x, lo, hi)` | Clamp `x` into a range |
 | `step(edge, x)` | `0` below the edge, `1` at or above it |
 | `smoothstep(edge0, edge1, x)` | Smooth curve between two edges |
diff --git a/src/checker.rs b/src/checker.rs
index 0ebd8e0d..220c0bdd 100644
--- a/src/checker.rs
+++ b/src/checker.rs
@@ -23,6 +23,11 @@ pub struct Checker {
     /// Overloads for arithmetic with built-in types.
     arith_overloads: Vec<TypeID>,
 
+    /// Overloads for `%` with built-in types. Integers only: float modulo is
+    /// provided by the `__mod` overloads in the stdlib, since no backend has a
+    /// primitive float remainder instruction.
+    mod_overloads: Vec<TypeID>,
+
     /// Overloads for casting.
     cast_overloads: Vec<TypeID>,
 
@@ -154,14 +159,19 @@ impl Checker {
         let uint32: TypeID = mk_type(Type::UInt32);
         let types = [Type::Int32, Type::UInt32, Type::Float32, Type::Float64];
         let mut arith_overloads = vec![];
+        let mut mod_overloads = vec![];
         let mut rel_overloads = vec![];
         let mut neg_overloads = vec![];
         let b = mk_type(Type::Bool);
 
         for ty in types {
+            let is_int = matches!(ty, Type::Int32 | Type::UInt32);
             let t = mk_type(ty);
             let tt = tuple(vec![t, t]);
             arith_overloads.push(func(tt, t));
+            if is_int {
+                mod_overloads.push(func(tt, t));
+            }
             rel_overloads.push(func(tt, b));
             neg_overloads.push(func(t, t));
         }
@@ -196,6 +206,7 @@ impl Checker {
             next_anon: 0,
             vars: vec![],
             arith_overloads,
+            mod_overloads,
             rel_overloads,
             neg_overloads,
             cast_overloads,
@@ -352,7 +363,13 @@ impl Checker {
 
             let mut alts = vec![];
 
-            for ty in &self.arith_overloads {
+            let builtins = if op == Binop::Mod {
+                &self.mod_overloads
+            } else {
+                &self.arith_overloads
+            };
+
+            for ty in builtins {
                 alts.push(Alt {
                     ty: *ty,
                     interfaces: vec![],
diff --git a/src/compiler.rs b/src/compiler.rs
index 6cbaf2e2..ff97305d 100644
--- a/src/compiler.rs
+++ b/src/compiler.rs
@@ -302,6 +302,9 @@ fn rewrite_qualified_enums(arena: &mut ExprArena, decls: &DeclTable) {
 /// `Call(__add/__sub/__mul/__div, [lhs, rhs])` when the operand type is a
 /// named (struct) type. The checker resolves the overload through its Or
 /// constraint, but the JIT/VM only handle primitive types in binop codegen.
+///
+/// Float `%` is rewritten the same way: no backend has a primitive float
+/// remainder instruction, so it lowers to the stdlib's `__mod` overloads.
 fn rewrite_overloaded_binops(fdecl: &mut FuncDecl) {
     let n = fdecl.arena.exprs.len();
     for i in 0..n {
@@ -310,7 +313,8 @@ fn rewrite_overloaded_binops(fdecl: &mut FuncDecl) {
                 continue;
             }
             let lhs_ty = fdecl.types[lhs];
-            if matches!(*lhs_ty, Type::Name(_, _)) {
+            let is_float_mod = op == Binop::Mod && matches!(*lhs_ty, Type::Float32 | Type::Float64);
+            if matches!(*lhs_ty, Type::Name(_, _)) || is_float_mod {
                 let result_ty = fdecl.types[i];
                 let rhs_ty = fdecl.types[rhs];
                 // Build the function type: (lhs_ty, rhs_ty) -> result_ty
diff --git a/src/llvm_jit.rs b/src/llvm_jit.rs
index d984f1d9..dd86900e 100644
--- a/src/llvm_jit.rs
+++ b/src/llvm_jit.rs
@@ -2642,6 +2642,11 @@ impl<'a, 'ctx> FunctionTranslator<'a, 'ctx> {
                 let lhs = self.translate_expr(lhs_id, decl);
                 let rhs = self.translate_expr(rhs_id, decl);
                 match *t {
+                    // Float `%` is lowered to a call to the stdlib's `__mod`
+                    // before codegen, so only integer operands reach here.
+                    crate::Type::Float32 | crate::Type::Float64 | crate::Type::Float32x4 => {
+                        unreachable!("type {:?} not supported for modulo", t)
+                    }
                     crate::Type::Int32 | crate::Type::Int8 => self
                         .builder()
                         .build_int_signed_rem(lhs.into_int_value(), rhs.into_int_value(), "srem")
diff --git a/src/stack_codegen.rs b/src/stack_codegen.rs
index 1f76be5a..f3e4bf36 100644
--- a/src/stack_codegen.rs
+++ b/src/stack_codegen.rs
@@ -1205,7 +1205,14 @@ impl<'a> FunctionTranslator<'a> {
                 Type::UInt32 | Type::UInt8 => func.emit(StackOp::UDiv),
                 _ => func.emit(StackOp::IDiv),
             },
-            Binop::Mod => func.emit(StackOp::IRem),
+            // Float `%` is lowered to a call to the stdlib's `__mod` before
+            // codegen, so only integer operands reach here.
+            Binop::Mod => match &*ty {
+                Type::Float32 | Type::Float64 | Type::Float32x4 => {
+                    unreachable!("type {:?} not supported for modulo", ty)
+                }
+                _ => func.emit(StackOp::IRem),
+            },
             Binop::Pow => match &*ty {
                 Type::Float32 => func.emit(StackOp::FPowF),
                 Type::Float64 => func.emit(StackOp::DPowD),
diff --git a/src/vm_codegen.rs b/src/vm_codegen.rs
index aa1b88f3..23e173a7 100644
--- a/src/vm_codegen.rs
+++ b/src/vm_codegen.rs
@@ -1728,13 +1728,20 @@ impl<'a> FunctionTranslator<'a> {
                 }
             },
 
-            Binop::Mod => {
-                func.emit(Opcode::IRem {
-                    dst,
-                    a: lhs,
-                    b: rhs,
-                });
-            }
+            // Float `%` is lowered to a call to the stdlib's `__mod` before
+            // codegen, so only integer operands reach here.
+            Binop::Mod => match &*ty {
+                Type::Float32 | Type::Float64 | Type::Float32x4 => {
+                    unreachable!("type {:?} not supported for modulo", ty)
+                }
+                _ => {
+                    func.emit(Opcode::IRem {
+                        dst,
+                        a: lhs,
+                        b: rhs,
+                    });
+                }
+            },
 
             Binop::Equal => match &*ty {
                 Type::Float32 => {
diff --git a/stdlib.lyte b/stdlib.lyte
index 442718be..f6fb770d 100644
--- a/stdlib.lyte
+++ b/stdlib.lyte
@@ -8,6 +8,25 @@ fract(x: f64) -> f64 { x - floor(x) }
 mod(x: f32, y: f32) -> f32 { x - y * floor(x / y) }
 mod(x: f64, y: f64) -> f64 { x - y * floor(x / y) }
 
+// Float `%`. No backend has a primitive float remainder instruction, so the
+// checker resolves `x % y` on floats to these overloads. Semantics match C's
+// fmod and Rust's `%`: the result takes the sign of x. Use mod() above for
+// floored modulo, whose result takes the sign of y. Unlike a true fmod, this
+// loses precision when x / y is large enough to round.
+__mod(x: f32, y: f32) -> f32 {
+    let q = x / y
+    var t = floor(q)
+    if q < 0.0 { t = ceil(q) }
+    x - y * t
+}
+
+__mod(x: f64, y: f64) -> f64 {
+    let q = x / y
+    var t = floor(q)
+    if q < 0.0f64 { t = ceil(q) }
+    x - y * t
+}
+
 clamp(x: f32, lo: f32, hi: f32) -> f32 { min(max(x, lo), hi) }
 clamp(x: f64, lo: f64, hi: f64) -> f64 { min(max(x, lo), hi) }
 
diff --git a/tests/cases/arith/float_modulo.lyte b/tests/cases/arith/float_modulo.lyte
new file mode 100644
index 00000000..06463bb5
--- /dev/null
+++ b/tests/cases/arith/float_modulo.lyte
@@ -0,0 +1,39 @@
+// expected stdout:
+// compilation successful
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+// assert(true)
+
+near(a: f32, b: f32) -> bool { abs(a - b) < 0.0001 }
+near64(a: f64, b: f64) -> bool { abs(a - b) < 0.0001f64 }
+
+// Wrap a value into [0, 1), the DSP idiom from issue #31.
+wrap01(x_in: f32) -> f32 {
+    var x = x_in % 1.0
+    if x < 0.0 { x = x + 1.0 }
+    x
+}
+
+main {
+    // `%` on floats is a truncated remainder (C fmod / Rust `%`):
+    // the result takes the sign of the left operand.
+    assert(near(5.5 % 2.0, 1.5))
+    assert(near(-5.5 % 2.0, -1.5))
+    assert(near(5.5 % -2.0, 1.5))
+    assert(near(4.0 % 2.0, 0.0))
+
+    assert(near64(5.5f64 % 2.0f64, 1.5f64))
+    assert(near64(-5.5f64 % 2.0f64, -1.5f64))
+
+    assert(near(wrap01(2.25), 0.25))
+    assert(near(wrap01(-0.25), 0.75))
+
+    // Integer `%` is unchanged.
+    assert(10 % 3 == 1)
+}