diff --git a/src/ggml-bitnet-mad.cpp b/src/ggml-bitnet-mad.cpp
index 4ba9d65..f16e3be 100644
--- a/src/ggml-bitnet-mad.cpp
+++ b/src/ggml-bitnet-mad.cpp
@@ -1043,2 +1043,25 @@ void ggml_vec_dot_i2_i8_s_Nx1(int n, float * s, size_t bs, const void * vx, size
 void ggml_vec_dot_i2_i8_s(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
+#if defined(__ARM_NEON)
+    assert(n % 128 == 0);
+    (void)bx;
+    const auto * packed = static_cast<const uint8_t *>(vx);
+    for (int col = 0; col < nrc; ++col) {
+        const auto * activation = static_cast<const int8_t *>(vy) + col * by;
+        int32x4_t sum = vdupq_n_s32(0);
+        for (int block = 0; block < n / 128; ++block) {
+            for (int group = 0; group < 4; ++group) {
+                for (int half = 0; half < 32; half += 16) {
+                    const uint8x16_t bytes = vld1q_u8(packed + block * 32 + half);
+                    const uint8x16_t codes = vandq_u8(
+                        vshlq_u8(bytes, vdupq_n_s8(2 * group - 6)), vdupq_n_u8(3));
+                    const int8x16_t values = vld1q_s8(activation + block * 128 + group * 32 + half);
+                    const int8x16_t weights = vreinterpretq_s8_u8(codes);
+                    sum = vpadalq_s16(sum, vmull_s8(vget_low_s8(weights), vget_low_s8(values)));
+                    sum = vpadalq_s16(sum, vmull_s8(vget_high_s8(weights), vget_high_s8(values)));
+                }
+            }
+        }
+        s[col * bs] = static_cast<float>(vaddvq_s32(sum));
+    }
+#else
     if (nrc % PARALLEL_SIZE == 0)
@@ -1055,2 +1078,3 @@ void ggml_vec_dot_i2_i8_s(int n, float * s, size_t bs, const void * vx, size_t b
     }
+#endif
 }
\ No newline at end of file