diff --git a/ggml/src/ggml-metal/ggml-metal-device.cpp b/ggml/src/ggml-metal/ggml-metal-device.cpp
index 3420ba85116c..260fdb431d9e 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-device.cpp
@@ -844,7 +844,17 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm(ggml_meta
     const int16_t r2   = (int16_t) (ne12 / op->src[0]->ne[2]);
     const int16_t r3   = (int16_t) (ne13 / op->src[0]->ne[3]);
 
-    snprintf(base, 256, "kernel_mul_mm_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
+    // Q1_0 products made only of full tiles can use the static-K32 tensor kernel (GGML_METAL_Q1_MM_K32=1)
+    static const bool q1_k32_env = getenv("GGML_METAL_Q1_MM_K32") && atoi(getenv("GGML_METAL_Q1_MM_K32")) == 1;
+    const bool q1_k32 = q1_k32_env && has_tensor && tsrc0 == GGML_TYPE_Q1_0 && tsrc1 == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                        !bc_inp && !bc_out && op->ne[0] >= NRA && op->ne[1] >= NRB &&
+                        ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && ggml_is_contiguous(op);
+
+    if (q1_k32) {
+        snprintf(base, 256, "kernel_mul_mm_q1_0_f32_k32");
+    } else {
+        snprintf(base, 256, "kernel_mul_mm_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
+    }
     snprintf(name, 256, "%s_bci=%d_bco=%d_ne12=%d_ne13=%d_r2=%d_r3=%d",
              base, bc_inp, bc_out, ne12, ne13, r2, r3);
 
diff --git a/ggml/src/ggml-metal/kernels/mul_mm.metal b/ggml/src/ggml-metal/kernels/mul_mm.metal
index b0029fdc6ad2..106dcea32547 100644
--- a/ggml/src/ggml-metal/kernels/mul_mm.metal
+++ b/ggml/src/ggml-metal/kernels/mul_mm.metal
@@ -140,6 +140,83 @@ kernel void kernel_mul_mm(
     cT.store(tD.slice(ra, rb));
 }
 
+// Q1_0 products made only of full tiles (M % 64, N % 128, K % 32 all zero): kernel_mul_mm with a static K32 extent and no bounds handling
+kernel void kernel_mul_mm_q1_0_f32_k32(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+    constexpr int NK  = N_MM_NK_TOTAL;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+    static_assert(NUM_THREADS == NRA * N_MM_NK, "one 16-weight chunk per thread");
+
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    const int im  = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // same work mapping as kernel_mul_mm: row = tiitg / N_MM_NK, chunk = tiitg % N_MM_NK
+    const int   row    = tiitg / N_MM_NK;
+    const short k_base = (tiitg % N_MM_NK) * 16;
+
+    threadgroup half * sa = (threadgroup half *) shmem;
+
+    device const block_q1_0 * row_ptr = (device const block_q1_0 *)(srcA + args.nb01 * (ra + row) + offset0);
+    device float * ptrB = (device float *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(float);
+
+    auto tA = tensor(sa, dextents<int32_t, 2>(NK, NRA));
+    auto tB = tensor(ptrB + rb * strideB, dextents<int32_t, 2>(NK, NRB), array<int, 2>({1, strideB}));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, NK, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    for (int loop_k = 0; loop_k < K; loop_k += NK) {
+        const int k_pos = loop_k + k_base;
+
+        half4x4 temp_a;
+        dequantize_q1_0(row_ptr + k_pos / QK1_0, (k_pos / 16) % (QK1_0 / 16), temp_a);
+
+        FOR_UNROLL (short i = 0; i < 16; i++) {
+            sa[row * NK + k_base + i] = temp_a[i/4][i%4];
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(NK, NRB), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device float * dstTile = (device float *)dst + (uint64_t) im * N * M + (uint64_t) rb * M + ra;
+
+    auto tD = tensor(dstTile, dextents<int32_t, 2>(NRA, NRB), array<int, 2>({1, M}));
+    cT.store(tD);
+}
+
 #else
 
 template<
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 72e580d895c0..9841980e06a5 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -9400,6 +9400,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    // Q1_0 products made only of full tensor tiles (GGML_METAL_Q1_MM_K32)
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q1_0, GGML_TYPE_F32, 128, 256, 512, {1, 1}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q1_0, GGML_TYPE_F32, 64, 128, 256, {2, 3}, {2, 1}));
+
     // PTQ1_0 small batches, row tails and broadcast dimensions.
     for (int n : {1, 2, 3, 4, 8}) {
         for (int k : {128, 384, 5120}) {