diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp
index 3420ba85116c..cbcdb36c1f6a 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-device.cpp
@@ -882,11 +882,21 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm(ggml_meta
     return res;
 }
 
+// widest batch the multi-column kernel takes, as column tiles of at most four; wider batches keep the default route.
+// GGML_METAL_PTQ1_MULTICOL_MAX is clamped to 4..8 (default 8); 4 keeps the kernel to 2-4 columns
+static int ggml_metal_ptq1_multicol_max(void) {
+    static const int max_cols = [] {
+        const char * env = getenv("GGML_METAL_PTQ1_MULTICOL_MAX");
+        return env ? std::min(8, std::max(4, atoi(env))) : 8;
+    }();
+    return max_cols;
+}
+
 bool ggml_metal_ptq1_multicol_enabled(const ggml_tensor * op) {
     static const bool enabled = getenv("GGML_METAL_PTQ1_MULTICOL") && atoi(getenv("GGML_METAL_PTQ1_MULTICOL")) == 1;
     return enabled && op->src[0]->type == GGML_TYPE_PTQ1_0 && op->src[1]->type == GGML_TYPE_F32 &&
            op->src[0]->ne[0] % ggml_blck_size(GGML_TYPE_PTQ1_0) == 0 && op->src[1]->nb[0] == sizeof(float) &&
-           op->src[1]->ne[1] >= 2 && op->src[1]->ne[1] <= 4;
+           op->src[1]->ne[1] >= 2 && op->src[1]->ne[1] <= ggml_metal_ptq1_multicol_max();
 }
 
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_metal_library_t lib, const ggml_tensor * op) {
@@ -907,6 +917,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
 
     const char * suffix = "";
     char ptq1_suffix[16];
+    bool ptq1_mc = false;
 
     // use custom matrix x vector kernel
     switch (tsrc0) {
@@ -959,9 +970,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
                 nsg = N_SG_PTQ1_0;
                 nr0 = N_R0_PTQ1_0;
                 if (ggml_metal_ptq1_multicol_enabled(op)) {
+                    ptq1_mc = true;
                     nr0 = 4;
                     nsg = 1;
-                    nr1 = ne11;
+                    // at most four columns per tile, split evenly: 5 -> 3+2, 6 -> 3+3, 7 -> 4+3, 8 -> 4+4
+                    nr1 = (ne11 + (ne11 + 3)/4 - 1) / ((ne11 + 3)/4);
                     snprintf(ptq1_suffix, sizeof(ptq1_suffix), "_mc_c%d", nr1);
                     suffix = ptq1_suffix;
                 }
@@ -1091,7 +1104,13 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
     const int16_t r3 = (int16_t) (ne13 / ne03);
 
     snprintf(base, 256, "kernel_mul_mv_%s_%s%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1), suffix);
-    snprintf(name, 256, "%s_nsg=%d_ne12=%d_r2=%d_r3=%d", base, nsg, ne12, r2, r3);
+    // PTQ1_0 multi-column: every column tile complete (2-4, 6, 8 columns) skips the partial-tile checks
+    const bool ptq1_full_cols = ptq1_mc && ne11 % nr1 == 0;
+    if (ptq1_mc) {
+        snprintf(name, 256, "%s_nsg=%d_ne12=%d_r2=%d_r3=%d_full=%d", base, nsg, ne12, r2, r3, ptq1_full_cols);
+    } else {
+        snprintf(name, 256, "%s_nsg=%d_ne12=%d_r2=%d_r3=%d", base, nsg, ne12, r2, r3);
+    }
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
     if (!res.pipeline) {
@@ -1101,6 +1120,9 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
         ggml_metal_cv_set_int16(cv, (int16_t) ne12, FC_MUL_MV + 2);
         ggml_metal_cv_set_int16(cv, r2,             FC_MUL_MV + 3);
         ggml_metal_cv_set_int16(cv, r3,             FC_MUL_MV + 4);
+        if (ptq1_mc) {
+            ggml_metal_cv_set_bool(cv, ptq1_full_cols, FC_MUL_MV + 5);
+        }
 
         res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
 
diff --git a/ggml/src/ggml-metal/kernels/mul_mv.metal b/ggml/src/ggml-metal/kernels/mul_mv.metal
index 7d3bbc0bee89..b3d6ce7c17b8 100644
--- a/ggml/src/ggml-metal/kernels/mul_mv.metal
+++ b/ggml/src/ggml-metal/kernels/mul_mv.metal
@@ -262,6 +262,7 @@ constant short FC_mul_mv_nxpsg [[function_constant(FC_MUL_MV + 1)]];
 constant short FC_mul_mv_ne12  [[function_constant(FC_MUL_MV + 2)]];
 constant short FC_mul_mv_r2    [[function_constant(FC_MUL_MV + 3)]];
 constant short FC_mul_mv_r3    [[function_constant(FC_MUL_MV + 4)]];
+constant bool  FC_mul_mv_ptq1_full_cols [[function_constant(FC_MUL_MV + 5)]];
 
 template<typename block_q_type, short NR0, typename args_t>
 void mul_vec_q_n_f32_impl(
@@ -1061,6 +1062,9 @@ kernel void kernel_mul_mv_ptq1_0_multicol(
     const int r1 = tgpig.y * nr1;
     const int im = tgpig.z;
 
+    // columns of this tile that exist: nr1 except in a partial last tile (5 or 7 columns)
+    const short ncols = FC_mul_mv_ptq1_full_cols ? nr1 : (short) min(nr1, args.ne11 - r1);
+
     const int first_row = (r0 * NSG + sgitg) * nr0;
 
     const uint i12 = im%FC_mul_mv_ne12;
@@ -1097,7 +1101,9 @@ kernel void kernel_mul_mv_ptq1_0_multicol(
         // Reuse collapse coefficients across rows: c[k-1] = y_{k-1} - 3*y_k, c[4] = y_4.
         float sumy[nr1] = {};
         FOR_UNROLL (short col = 0; col < nr1; ++col) {
-            device const float * yc = (device const float *) ((device const char *) yb + col*args.nb11);
+            // a partial last tile re-reads its final valid column; that result is not written
+            const short yc_col = FC_mul_mv_ptq1_full_cols ? col : min(col, (short) (ncols - 1));
+            device const float * yc = (device const float *) ((device const char *) yb + yc_col*args.nb11);
 
             FOR_UNROLL (short k = 0; k < 2; ++k) {
                 const short m = 2*it + k;
@@ -1141,7 +1147,7 @@ kernel void kernel_mul_mv_ptq1_0_multicol(
     for (int row = 0; row < nr0; ++row) {
         FOR_UNROLL (short col = 0; col < nr1; ++col) {
             const float tot = simd_sum(sumf[row][col]);
-            if (tiisg == 0 && first_row + row < args.ne01) {
+            if (tiisg == 0 && first_row + row < args.ne01 && (FC_mul_mv_ptq1_full_cols || col < ncols)) {
                 dst_f32[(uint64_t) col*args.ne0 + first_row + row] = tot;
             }
         }
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 72e580d895c0..b6ac8ac988d3 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -9401,11 +9401,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     // PTQ1_0 small batches, row tails and broadcast dimensions.
-    for (int n : {1, 2, 3, 4, 8}) {
+    for (int n : {1, 2, 3, 4, 5, 6, 7, 8}) {
         for (int k : {128, 384, 5120}) {
             test_cases.emplace_back(new test_mul_mat(GGML_TYPE_PTQ1_0, GGML_TYPE_F32, 7, n, k, {2, 2}, {2, 1}));
         }
     }
+    // partial column tiles with a strided B
+    for (int n : {5, 7}) {
+        test_cases.emplace_back(new test_mul_mat(GGML_TYPE_PTQ1_0, GGML_TYPE_F32, 7, n, 384, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+    }
 
     // BF16 is absent from base_types: add the 3 standard non-contig permutations explicitly
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_BF16, GGML_TYPE_F32, 16,  1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));