diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index cd39c2a..225b49f 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -10,8 +10,11 @@ env:
 
 jobs:
   rust:
-    name: Rust tests
-    runs-on: ubuntu-latest
+    name: Rust tests (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-14]
     steps:
       - uses: actions/checkout@v4
 
@@ -27,9 +30,9 @@ jobs:
             ~/.cargo/registry
             ~/.cargo/git
             target
-          key: cargo-${{ runner.os }}-${{ hashFiles('Cargo.toml') }}
+          key: cargo-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('Cargo.toml') }}
           restore-keys: |
-            cargo-${{ runner.os }}-
+            cargo-${{ runner.os }}-${{ runner.arch }}-
 
       - name: Build (release)
         run: cargo build --release
@@ -37,7 +40,7 @@ jobs:
       - name: Test (release)
         run: cargo test --release
 
-      - name: Test scalar-only (no AVX2 feature)
+      - name: Test scalar-only (no AVX2/NEON feature)
         run: cargo test --release --no-default-features
 
       - name: Test with test-panic feature
diff --git a/src/scan/neon.rs b/src/scan/neon.rs
index 7c1db0a..2242da3 100644
--- a/src/scan/neon.rs
+++ b/src/scan/neon.rs
@@ -88,25 +88,6 @@ unsafe fn tag_mask16(tag: uint8x16_t, bits: u8) -> u16 {
     movemask16(vtstq_u8(tag, vdupq_n_u8(bits)))
 }
 
-#[inline(always)]
-unsafe fn byte_mask16(bytes: uint8x16_t, needle: u8) -> u16 {
-    movemask16(vceqq_u8(bytes, vdupq_n_u8(needle)))
-}
-
-#[inline(always)]
-unsafe fn byte_mask64(
-    c0: uint8x16_t,
-    c1: uint8x16_t,
-    c2: uint8x16_t,
-    c3: uint8x16_t,
-    needle: u8,
-) -> u64 {
-    (byte_mask16(c0, needle) as u64)
-        | ((byte_mask16(c1, needle) as u64) << 16)
-        | ((byte_mask16(c2, needle) as u64) << 32)
-        | ((byte_mask16(c3, needle) as u64) << 48)
-}
-
 #[inline(always)]
 unsafe fn classify_tags64(
     c0: uint8x16_t,
@@ -170,10 +151,19 @@ unsafe fn scan_neon_impl(buf: &[u8], out: &mut Vec<u32>) -> Result<(), usize> {
 
         // In-string fast probe: while already in a string, avoid the full
         // nibble-LUT classification unless this block contains quote/backslash.
+        // Uses vmaxvq_u8 on OR'd comparison results instead of byte_mask64 to
+        // avoid the expensive movemask16 pairwise-add chain (~3x faster probe).
         if in_string != 0 {
-            let quote_probe = byte_mask64(c0, c1, c2, c3, b'"');
-            let backslash_probe = byte_mask64(c0, c1, c2, c3, b'\\');
-            if (quote_probe | backslash_probe) == 0 {
+            let quote = vdupq_n_u8(b'"');
+            let backslash = vdupq_n_u8(b'\\');
+            let m0 = vorrq_u8(vceqq_u8(c0, quote), vceqq_u8(c0, backslash));
+            let m1 = vorrq_u8(vceqq_u8(c1, quote), vceqq_u8(c1, backslash));
+            let m2 = vorrq_u8(vceqq_u8(c2, quote), vceqq_u8(c2, backslash));
+            let m3 = vorrq_u8(vceqq_u8(c3, quote), vceqq_u8(c3, backslash));
+            let m01 = vorrq_u8(m0, m1);
+            let m23 = vorrq_u8(m2, m3);
+            let m = vorrq_u8(m01, m23);
+            if vmaxvq_u8(m) == 0 {
                 bs_carry = 0;
                 i += 64;
                 // Cross-chunk jump: with no quote/backslash in the chunk we just