diff --git a/crates/jam-sema/src/astgen.rs b/crates/jam-sema/src/astgen.rs
index c0b7d34..581703e 100644
--- a/crates/jam-sema/src/astgen.rs
+++ b/crates/jam-sema/src/astgen.rs
@@ -7981,6 +7981,13 @@ fn astgen_for(gctx: &mut AstGenCtx, n: &AstNode) -> Result<(), String> {
                 start_hint = t;
             }
         }
+        // A literal end bound makes the loop var u64. It used to type from
+        // the START literal (`0` -> u8), so `for i in 0:256` truncated the
+        // bound through u8 and never ran; smallest-fit on the end is no
+        // better (`i * i` wraps at u16 for `0:1024`). Indices are u64.
+        if en.tag == AstTag::NumberLit && en.flags & 3 == 0 {
+            start_hint = builtin::U64;
+        }
     }
     let start_ref = astgen_expr(gctx, start_idx, start_hint)?;
     let idx_ty = gctx.jfn.get_inst(start_ref).ty;
diff --git a/tests/unit/test_for_range_bounds.jam b/tests/unit/test_for_range_bounds.jam
new file mode 100644
index 0000000..08d6de9
--- /dev/null
+++ b/tests/unit/test_for_range_bounds.jam
@@ -0,0 +1,28 @@
+// The loop index takes its width from a literal end bound. It used to
+// type from the start literal (`0` -> u8), so `for i in 0:256` truncated
+// the bound through u8 and never ran.
+const { assert } = import("test");
+
+tfn forBoundPastU8() {
+    var count: u64 = 0;
+    for i in 0:256 {
+        count = count + 1;
+    }
+    assert(count as i32, 256);
+}
+
+tfn forBoundPastU16() {
+    var count: u64 = 0;
+    for i in 0:70000 {
+        count = count + 1;
+    }
+    assert((count == 70000) as u8, 1);
+}
+
+tfn forSmallBoundStillWorks() {
+    var sum: u64 = 0;
+    for i in 0:10 {
+        sum = sum + i;
+    }
+    assert(sum as i32, 45);
+}