diff --git a/javascript/packages/core/lib/types/float16.ts b/javascript/packages/core/lib/types/float16.ts
index de2fc9100d..20ffcd332c 100644
--- a/javascript/packages/core/lib/types/float16.ts
+++ b/javascript/packages/core/lib/types/float16.ts
@@ -35,6 +35,13 @@ export function toFloat16Bits(value: number) {
     return sign | 0x7c00;
   }
 
+  if (exponent < -24) {
+    // Too small for a float16 subnormal. Larger shifts below would wrap
+    // (JS masks shift counts with & 31) and leave garbage bits, so
+    // underflow to signed zero.
+    return sign;
+  }
+
   if (exponent < -14) {
     return sign | ((significand | 0x800000) >> (13 - 14 - exponent));
   }
diff --git a/javascript/test/number.test.ts b/javascript/test/number.test.ts
index c1e84754f8..7da913422c 100644
--- a/javascript/test/number.test.ts
+++ b/javascript/test/number.test.ts
@@ -154,6 +154,20 @@ describe("number", () => {
     expect(result.a).toBe(NaN);
   });
 
+  test("should float16 underflow tiny magnitudes to signed zero", () => {
+    // Magnitudes below the smallest float16 subnormal must encode as zero;
+    // shift counts of 32 or more wrapped (JS masks them with & 31) and left
+    // garbage bits in the half.
+    const fory = new Fory({ compatible: false, ref: true });
+    const { serialize, deserialize } = fory.register(
+      Type.struct({ typeName: "example.f16zero" }, { a: Type.float16() }),
+    );
+    expect(deserialize(serialize({ a: 1e-10 })).a).toBe(0);
+    expect(deserialize(serialize({ a: 1e-11 })).a).toBe(0);
+    expect(deserialize(serialize({ a: 1e-40 })).a).toBe(0);
+    expect(deserialize(serialize({ a: -1e-10 })).a).toBe(-0);
+  });
+
   test("should float16 Infinity work", () => {
     const fory = new Fory({ compatible: false, ref: true });
     const serializer = fory.register(