diff --git a/ggml/src/iqk/iqk_gemm_kquants.cpp b/ggml/src/iqk/iqk_gemm_kquants.cpp
index db732383..05baea14 100644
--- a/ggml/src/iqk/iqk_gemm_kquants.cpp
+++ b/ggml/src/iqk/iqk_gemm_kquants.cpp
@@ -2064,14 +2064,21 @@ typedef struct {
 } block_q8_1_r8;
 
 void iqk_convert_q2_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
+#ifdef HAVE_FANCY_SIMD
+    constexpr int k_nr = 16;
+    using block_q8_k_r = block_q8_k_r16;
+#else
+    constexpr int k_nr = 8;
+    using block_q8_k_r = block_q8_k_r8;
+#endif
     GGML_ASSERT(n%QK_K == 0);
-    GGML_ASSERT(nrc_x%8 == 0);
+    GGML_ASSERT(nrc_x%k_nr == 0);
 
     int nb = n/QK_K;
 
-    const block_q2_K * x8[8];
+    const block_q2_K * x8[k_nr];
 
-    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+    block_q8_k_r * y = (block_q8_k_r *)vy;
 
     float    f_values[QK_K];
     uint32_t block[8];
@@ -2082,10 +2089,10 @@ void iqk_convert_q2_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
     auto sign_bit = _mm256_set1_ps(-0.0f);
     auto perm = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7);
 
-    for (int ix = 0; ix < nrc_x; ix += 8) {
-        for (int k = 0; k < 8; ++k) x8[k] = (const block_q2_K *)((const char *)vx + (ix + k)*bx);
+    for (int ix = 0; ix < nrc_x; ix += k_nr) {
+        for (int k = 0; k < k_nr; ++k) x8[k] = (const block_q2_K *)((const char *)vx + (ix + k)*bx);
         for (int i = 0; i < nb; ++i) {
-            for (int k = 0; k < 8; ++k) {
+            for (int k = 0; k < k_nr; ++k) {
                 auto vd = _mm256_set1_ps(GGML_FP16_TO_FP32(x8[k][i].d));
                 auto vm = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x8[k][i].dmin)), _mm256_set1_ps(-1.f));
                 auto block_max = _mm256_setzero_ps();
@@ -2136,13 +2143,18 @@ void iqk_convert_q2_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
                     i0 = _mm256_permutevar8x32_epi32(i0, perm);
 
                     _mm256_storeu_si256((__m256i *)block, i0);
-                    auto q8 = (uint32_t *)y[i].qs + 64*ib32;
-                    for (int l = 0; l < 4; ++l) {
-                        q8[8*l + k +  0] = block[l + 0];
-                        q8[8*l + k + 32] = block[l + 4];
+                    auto q8 = (uint32_t *)y[i].qs + 8*k_nr*ib32;
+                    for (int l = 0; l < 8; ++l) {
+                        q8[k_nr*l + k] = block[l];
                     }
                 }
             }
+#ifdef HAVE_FANCY_SIMD
+            for (int l = 0; l < 64; ++l) {
+                auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
+                _mm512_storeu_si512((__m512i *)y[i].qs + l, v);
+            }
+#endif
         }
         y += nb;
     }
@@ -2458,14 +2470,21 @@ void iqk_convert_q3_k_q8_0_r8(int n, const void * vx, size_t bx, void * vy, int
 }
 
 void iqk_convert_q3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
+#ifdef HAVE_FANCY_SIMD
+    constexpr int k_nr = 16;
+    using block_q8_k_r = block_q8_k_r16;
+#else
+    constexpr int k_nr = 8;
+    using block_q8_k_r = block_q8_k_r8;
+#endif
     GGML_ASSERT(n%QK_K == 0);
-    GGML_ASSERT(nrc_x%8 == 0);
+    GGML_ASSERT(nrc_x%k_nr == 0);
 
     int nb = n/QK_K;
 
-    const block_q3_K * x8[8];
+    const block_q3_K * x8[k_nr];
 
-    block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
+    block_q8_k_r * y = (block_q8_k_r *)vy;
 
     uint32_t block[8];
     __m256i values[8];
@@ -2476,10 +2495,10 @@ void iqk_convert_q3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
 
     union { __m256i vec; int16_t val[16]; } helper;
 
-    for (int ix = 0; ix < nrc_x; ix += 8) {
-        for (int k = 0; k < 8; ++k) x8[k] = (const block_q3_K *)((const char *)vx + (ix + k)*bx);
+    for (int ix = 0; ix < nrc_x; ix += k_nr) {
+        for (int k = 0; k < k_nr; ++k) x8[k] = (const block_q3_K *)((const char *)vx + (ix + k)*bx);
         for (int i = 0; i < nb; ++i) {
-            for (int k = 0; k < 8; ++k) {
+            for (int k = 0; k < k_nr; ++k) {
                 float d = GGML_FP16_TO_FP32(x8[k][i].d);
                 auto hbits = _mm256_loadu_si256((const __m256i *)x8[k][i].hmask);
                 helper.vec = _mm256_cvtepi8_epi16(sc3.make_scales((const uint16_t *)x8[k][i].scales));
@@ -2549,12 +2568,18 @@ void iqk_convert_q3_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
                         _mm_storeu_si128((__m128i *)block+0, _mm_unpacklo_epi64(i0_l, i0_h));
                         _mm_storeu_si128((__m128i *)block+1, _mm_unpackhi_epi64(i0_l, i0_h));
                     }
-                    auto qs = (uint32_t *)y[i].qs + 64*ib32;
+                    auto qs = (uint32_t *)y[i].qs + 8*k_nr*ib32;
                     for (int l = 0; l < 8; ++l) {
-                        qs[8*l + k] = block[l];
+                        qs[k_nr*l + k] = block[l];
                     }
                 }
             }
+#ifdef HAVE_FANCY_SIMD
+            for (int l = 0; l < 64; ++l) {
+                auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
+                _mm512_storeu_si512((__m512i *)y[i].qs + l, v);
+            }
+#endif
         }
         y += nb;
     }