diff --git a/mlx/backend/metal/kernels/quantized.h b/mlx/backend/metal/kernels/quantized.h
index 5e70fc8fe..d52380aed 100644
--- a/mlx/backend/metal/kernels/quantized.h
+++ b/mlx/backend/metal/kernels/quantized.h
@@ -2171,11 +2171,10 @@ inline vec<U, 4> partial_qdot_vec(const thread U* x, vec<uint32_t, 4> w) {
 
   else if (bits == 4) {
     for (int i = 0; i < 4; i++) {
-      auto ws = as_type<vec<uint16_t, 2>>(w[i]);
-      for (int j = 0; j < 2; j++) {
-        accum[i] +=
-            (x[4 * j + 0] * (ws[j] & 0x000f) + x[4 * j + 1] * (ws[j] & 0x00f0) +
-             x[4 * j + 2] * (ws[j] & 0x0f00) + x[4 * j + 3] * (ws[j] & 0xf000));
+      auto ws = as_type<vec<uint8_t, 4>>(w[i]);
+      for (int j = 0; j < 4; j++) {
+        accum[j] +=
+            x[2 * i + 0] * (ws[j] & 0x0f) + x[2 * i + 1] * (ws[j] & 0xf0);
       }
     }
   }
@@ -2184,7 +2183,7 @@ inline vec<U, 4> partial_qdot_vec(const thread U* x, vec<uint32_t, 4> w) {
     for (int i = 0; i < 4; i++) {
       auto ws = as_type<vec<uint8_t, 4>>(w[i]);
       for (int j = 0; j < 4; j++) {
-        accum[i] += x[j] * ws[j];
+        accum[j] += x[i] * ws[j];
       }
     }
   }
diff --git a/mlx/ops.cpp b/mlx/ops.cpp
index 7f6d296d1..6f4911fec 100644
--- a/mlx/ops.cpp
+++ b/mlx/ops.cpp
@@ -3795,9 +3795,11 @@ std::tuple<array, array, std::optional<array>> quantize(
       scales = moveaxis(scales, -2, -1, s);
       scales = flatten(scales, -2, -1, s);
 
+      wq = view(wq, uint8, s);
       wq = unflatten(wq, -2, {-1, 4}, s);
       wq = moveaxis(wq, -2, -1, s);
       wq = flatten(wq, -2, -1, s);
+      wq = view(wq, uint32, s);
 
       return std::make_tuple(wq, scales, std::nullopt);
     }