diff --git a/ggml-metal.metal b/ggml-metal.metal
index 43814ed09..981bb41af 100644
--- a/ggml-metal.metal
+++ b/ggml-metal.metal
@@ -267,6 +267,8 @@ kernel void kernel_mul_mat_q4_0_f32(
         uint2  tptg[[threads_per_threadgroup]]) {
     const int nb = ne00/QK4_0;
 
+    const int8_t m8 = 8;
+
     const int64_t r0 = tgpig.x;
     const int64_t r1 = tgpig.y;
 
@@ -276,33 +278,54 @@ kernel void kernel_mul_mat_q4_0_f32(
     const uint nth = tptg.x*tptg.y;
     const uint ith = tptg.y*tpitg.x + tpitg.y;
 
+    const int first = 4 * tpitg.y;
+
     sum[ith] = 0.0f;
 
+    float sumf = 0;
+    //float sumf1 = 0;
+    //float sumf2 = 0;
+
     for (int i = tpitg.x; i < nb; i += tptg.x) {
-        device const uchar4 * x0p = (device const uchar4 *) (x + i)->qs;
-        device const float4 * y0p = (device const float4 *) (y + i*QK4_0);
+        device const uchar * x0p = (device const uchar *) (x + i)->qs;
+        device const float * y0p = (device const float *) (y + i*QK4_0);
 
         const float d = (float)((x + i)->d);
 
-        const uchar4 x0v = *(x0p + tpitg.y);
-        const float4 y0v = *(y0p + tpitg.y + 0);
-        const float4 y1v = *(y0p + tpitg.y + 4);
+        device const uchar * x0v = x0p + first;
+        device const float * y0v = y0p + first;
+        device const float * y1v = y0p + first + 16;
 
-        float acc = 0.0f;
+        //float3 acc = {0.0f, 0.0f, 0.f};
+        float2 acc = {0.0f, 0.0f};
 
         for (int j = 0; j < 4; ++j) {
-            const int x0 = x0v[j] & 0x0F;
-            const int x1 = x0v[j] >>   4;
 
-            const float y0 = y0v[j];
-            const float y1 = y1v[j];
+            //acc[0] += y0v[j] * (x0v[j] & 0xF);
+            //acc[1] += y1v[j] * (x0v[j] >>  4);
+            //acc[2] += y0v[j] + y1v[j];
 
-            acc += (x0 - 8)*y0 + (x1 - 8)*y1;
+            acc[0] += y0v[j] * ((int8_t)(x0v[j] & 0xF) - m8);
+            acc[1] += y1v[j] * ((int8_t)(x0v[j] >>  4) - m8);
+
+            //const int x0 = x0v[j] & 0x0F;
+            //const int x1 = x0v[j] >>   4;
+
+            //const float y0 = y0v[j];
+            //const float y1 = y1v[j];
+
+            //acc += (x0 - 8)*y0 + (x1 - 8)*y1;
         }
 
-        sum[ith] += acc*d;
+        //sum[ith] += acc*d;
+        sumf += d * (acc[0] + acc[1]);
+        //sumf1 += d * (acc[0] + acc[1]);
+        //sumf2 += d * acc[2];
     }
 
+    sum[ith] = sumf;
+    //sum[ith] = sumf1 - 8.f*sumf2;
+
     //
     // Accumulate the sum from all threads in the threadgroup
     // This version is slightly faster than the commented out one below,