diff --git a/source/mir/parse.d b/source/mir/parse.d
index b870c213..2432c087 100644
--- a/source/mir/parse.d
+++ b/source/mir/parse.d
@@ -209,7 +209,43 @@ version(mir_bignum_test)
         else
         {
             str.fromString!float.should == str._stdc_parse!float;
-            str.fromString!double.should == str._stdc_parse!double;
+
+            // Some Windows CRTs clamp extreme underflow to min-subnormal instead of 0.
+            // Accept both behaviors *only* at the very bottom of the subnormal range.
+            bool equalOrTinyUnderflow(double a, double b)
+            {
+                import core.stdc.math: fpclassify, FP_SUBNORMAL;
+                import mir.math.common: fabs;
+                import mir.math.ieee: signbit;
+                import std.math.operations: nextafter;
+
+                if (a == b) return true;
+                if (signbit(a) != signbit(b)) return false;
+
+                enum double minSub = nextafter(0.0, 1.0); // smallest positive subnormal
+
+                const aa = fabs(a);
+                const bb = fabs(b);
+
+                // 0 <-> min-subnormal (or similar one-ULP-at-zero behavior)
+                if (aa == 0.0 && bb > 0.0)
+                    return fpclassify(b) == FP_SUBNORMAL && bb <= minSub;
+                if (bb == 0.0 && aa > 0.0)
+                    return fpclassify(a) == FP_SUBNORMAL && aa <= minSub;
+
+                // If both are subnormal, permit differences only if both are at the very bottom.
+                if (fpclassify(a) == FP_SUBNORMAL && fpclassify(b) == FP_SUBNORMAL)
+                    return aa <= minSub && bb <= minSub;
+
+                return false;
+            }
+
+            auto mirv = str.fromString!double;
+            auto crtv = str._stdc_parse!double;
+            if (!equalOrTinyUnderflow(mirv, crtv)) {
+                mirv.should == crtv;
+            }
+
             version (Windows) // No precise real parsing on windows
             {
             }