diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index 2859b5cb..247f7f65 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -6916,6 +6916,43 @@ void mul_mat_q4_0_r4_q8_0(int n, const void * vx, size_t bx, const DataInfo& inf
     }
 }
 
+template <int nrc_y>
+void mul_mat_q8_0_r4_q8_0(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+    GGML_ASSERT(nrc_x%4 == 0);
+    Q8<nrc_y, block_q8_0_x4> q8(info);
+    int nb = n / QK8_0;
+    GGML_ASSERT(nb%4 == 0);
+    float32x4_t acc[nrc_y] = {};
+    for (int ix = 0; ix < nrc_x; ix += 4) {
+        const block_q8_0_x4 * iq8 = (const block_q8_0_x4 *)((const char *)vx + ix*bx);
+        for (int ib4 = 0; ib4 < nb/4; ++ib4) {
+            for (int k = 0; k < 4; ++k) {
+                auto scales = vcvt_f32_f16(vld1_f16((const float16_t *)iq8[4*ib4+k].d));
+                auto qx1 = vld1q_s8_x4(iq8[4*ib4+k].qs);
+                auto qx2 = vld1q_s8_x4(iq8[4*ib4+k].qs+64);
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = vld1q_s8_x2(q8.y[iy][ib4].qs+32*k);
+                    auto sumi = vdupq_n_s32(0);
+                    sumi = vdotq_laneq_s32(sumi, qx1.val[0], y.val[0], 0);
+                    sumi = vdotq_laneq_s32(sumi, qx1.val[1], y.val[1], 0);
+                    sumi = vdotq_laneq_s32(sumi, qx1.val[2], y.val[0], 1);
+                    sumi = vdotq_laneq_s32(sumi, qx1.val[3], y.val[1], 1);
+                    sumi = vdotq_laneq_s32(sumi, qx2.val[0], y.val[0], 2);
+                    sumi = vdotq_laneq_s32(sumi, qx2.val[1], y.val[1], 2);
+                    sumi = vdotq_laneq_s32(sumi, qx2.val[2], y.val[0], 3);
+                    sumi = vdotq_laneq_s32(sumi, qx2.val[3], y.val[1], 3);
+                    auto d4d8 = vmulq_f32(scales, vdupq_n_f32(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k])));
+                    acc[iy] = vfmaq_f32(acc[iy], d4d8, vcvtq_f32_s32(sumi));
+                }
+            }
+        }
+        for (int iy = 0; iy < nrc_y; ++iy) {
+            info.store(ix, iy, acc[iy]);
+            acc[iy] = vdupq_n_f32(0.f);
+        }
+    }
+}
+
 template <typename Dequantizer> void MulMat::set_functions(MulMat& m) {
     if constexpr (std::is_same_v<Dequantizer, DequantizerQ40> || std::is_same_v<Dequantizer, DequantizerQ50> ||
                   std::is_same_v<Dequantizer, DequantizerQ80> || std::is_same_v<Dequantizer, DequantizerIQ4NL> ||
@@ -7107,6 +7144,17 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
             m.funcs[7] = mul_mat_q4_0_r4_q8_0<8>;
             expected_Btype = GGML_TYPE_Q8_0;
             break;
+        case GGML_TYPE_Q8_0_R4:
+            m.funcs[0] = mul_mat_q8_0_r4_q8_0<1>;
+            m.funcs[1] = mul_mat_q8_0_r4_q8_0<2>;
+            m.funcs[2] = mul_mat_q8_0_r4_q8_0<3>;
+            m.funcs[3] = mul_mat_q8_0_r4_q8_0<4>;
+            m.funcs[4] = mul_mat_q8_0_r4_q8_0<5>;
+            m.funcs[5] = mul_mat_q8_0_r4_q8_0<6>;
+            m.funcs[6] = mul_mat_q8_0_r4_q8_0<7>;
+            m.funcs[7] = mul_mat_q8_0_r4_q8_0<8>;
+            expected_Btype = GGML_TYPE_Q8_0;
+            break;
         default:
             return false;
     }