diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index fbb04426abe7e..bd4218d416dea 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -368,8 +368,8 @@ typedef struct {
 } block_iq3_xxs;
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
-// 3.4375 bpw
 #define IQ3S_N_SCALE QK_K/64
+// 3.4375 bpw
 typedef struct {
     ggml_half d;
     uint8_t qs[QK_K/4];
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index f4b5713d7dd9a..acdab324d2dec 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -84,10 +84,22 @@ struct quantize_state_impl {
     int n_ffn_down     = 0;
     int n_ffn_gate     = 0;
     int n_ffn_up       = 0;
+    int n_ffn_down_exps = 0;
+    int n_ffn_gate_exps = 0;
+    int n_ffn_up_exps   = 0;
+    int n_ffn_down_shexp = 0;
+    int n_ffn_gate_shexp = 0;
+    int n_ffn_up_shexp   = 0;
     int i_attention_wv = 0;
     int i_ffn_down     = 0;
     int i_ffn_gate     = 0;
     int i_ffn_up       = 0;
+    int i_ffn_down_exps = 0;
+    int i_ffn_gate_exps = 0;
+    int i_ffn_up_exps   = 0;
+    int i_ffn_down_shexp = 0;
+    int i_ffn_gate_shexp = 0;
+    int i_ffn_up_shexp   = 0;
 
     int n_k_quantized = 0;
     int n_fallback    = 0;
@@ -175,6 +187,56 @@ static void llama_tensor_dequantize_impl(
     workers.clear();
 }
 
+
+// Returns the appropriate type for expert _exps tensors based on ftype
+static inline ggml_type get_exps_type_low_bpw_bump(llama_ftype ftype, ggml_type new_type) {
+    if      (ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_XXS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S) new_type = GGML_TYPE_IQ2_S;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) new_type = GGML_TYPE_IQ2_S;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS) new_type = GGML_TYPE_IQ2_XS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) new_type = GGML_TYPE_IQ1_M;
+    return new_type;
+}
+
+static inline ggml_type get_exps_type_low_bpw_squash(llama_ftype ftype, ggml_type new_type) {
+    if      (ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ2_XS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S) new_type = GGML_TYPE_IQ2_XXS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) new_type = GGML_TYPE_IQ2_XXS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS) new_type = GGML_TYPE_IQ1_M;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ1_S;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) new_type = GGML_TYPE_IQ1_S;
+    return new_type;
+}
+
+static inline ggml_type get_exps_type_high_bpw_bump(llama_ftype ftype, ggml_type new_type, bool has_imatrix) {
+    if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S) new_type = GGML_TYPE_Q5_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q6_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
+    // Bump I-quants
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) new_type = GGML_TYPE_Q4_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
+    else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && !has_imatrix) new_type = GGML_TYPE_Q5_K;
+
+    return new_type;
+}
+
+static inline ggml_type get_exps_type_high_bpw_squash(llama_ftype ftype, ggml_type new_type, bool has_imatrix) {
+    // Squash K-quants
+    if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q2_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S) new_type = GGML_TYPE_Q3_K;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
+    // Squash I-quants
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) new_type = GGML_TYPE_IQ3_XXS;
+    else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
+        new_type = has_imatrix ? GGML_TYPE_IQ2_S : GGML_TYPE_Q2_K;
+    }
+    return new_type;
+}
+
 static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
     const std::string name = ggml_get_name(tensor);
 
@@ -245,21 +307,108 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M    || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
         if (name.find("attn_v.weight") != std::string::npos) {
             if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
-            else new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            else new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            ++qs.i_attention_wv;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("attn_k.weight") != std::string::npos) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("attn_kv_a_mqa.weight") != std::string::npos) {
+            new_type = GGML_TYPE_Q4_K;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("attn_kv_b.weight") != std::string::npos) {
+            if (qs.i_attention_wv < qs.n_attention_wv/8) {
+                new_type = GGML_TYPE_Q4_K;
+            }
+            else if (use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) {
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
             ++qs.i_attention_wv;
         }
-        else if (qs.model.hparams.n_expert == 8 && name.find("attn_k.weight") != std::string::npos) {
+        else if (qs.model.hparams.n_expert >= 8 && name.find("attn_q_a.weight") != std::string::npos) {
             new_type = GGML_TYPE_Q4_K;
         }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("attn_q_b.weight") != std::string::npos) {
+            new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_down.weight") != std::string::npos) {
+            if (qs.i_ffn_down < qs.n_ffn_down/16) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+            else if (qs.i_ffn_down < qs.n_ffn_down/8) {
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
+            ++qs.i_ffn_down;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_gate.weight") != std::string::npos) {
+            if (qs.i_ffn_gate < qs.n_ffn_gate/16) {
+                new_type = GGML_TYPE_Q4_K;
+            }
+            else if (qs.i_ffn_gate < qs.n_ffn_gate/8) {
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
+            ++qs.i_ffn_gate;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_up.weight") != std::string::npos) {
+            if (qs.i_ffn_up < qs.n_ffn_up/16) {
+                new_type = GGML_TYPE_Q4_K;
+            }
+            else if (qs.i_ffn_up < qs.n_ffn_up/8) {
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+            }
+            ++qs.i_ffn_up;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_down_exps.weight") != std::string::npos) {
+            if (qs.i_ffn_down_exps < qs.n_ffn_down_exps/8 || qs.i_ffn_down_exps > 7*qs.n_ffn_down_exps/8) {
+                new_type = get_exps_type_low_bpw_bump(ftype, new_type);
+            } else {
+                new_type = get_exps_type_low_bpw_squash(ftype, new_type);
+            }
+            ++qs.i_ffn_down_exps;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_gate_exps.weight") != std::string::npos) {
+            if (qs.i_ffn_gate_exps < qs.n_ffn_gate_exps/8 || qs.i_ffn_gate_exps > 7*qs.n_ffn_gate_exps/8) {
+                new_type = get_exps_type_low_bpw_bump(ftype, new_type);
+            } else {
+                new_type = get_exps_type_low_bpw_squash(ftype, new_type);
+            }
+            ++qs.i_ffn_gate_exps;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_up_exps.weight") != std::string::npos) {
+            if (qs.i_ffn_up_exps < qs.n_ffn_up_exps/8 || qs.i_ffn_up_exps > 7*qs.n_ffn_up_exps/8) {
+                new_type = get_exps_type_low_bpw_bump(ftype, new_type);
+            } else {
+                new_type = get_exps_type_low_bpw_squash(ftype, new_type);
+            }
+            ++qs.i_ffn_up_exps;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_down_shexp.weight") != std::string::npos) {
+            if (use_more_bits(qs.i_ffn_down_shexp, qs.n_ffn_down_shexp)) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+            ++qs.i_ffn_down_shexp;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_gate_shexp.weight") != std::string::npos) {
+            if (use_more_bits(qs.i_ffn_gate_shexp, qs.n_ffn_gate_shexp)) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+            ++qs.i_ffn_gate_shexp;
+        }
+        else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_up_shexp.weight") != std::string::npos) {
+            if (use_more_bits(qs.i_ffn_up_shexp, qs.n_ffn_up_shexp)) {
+                new_type = GGML_TYPE_Q6_K;
+            }
+            ++qs.i_ffn_up_shexp;
+        }
         else if (name.find("ffn_down") != std::string::npos) {
             if (qs.i_ffn_down < qs.n_ffn_down/8) {
-                new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K;
             }
             ++qs.i_ffn_down;
         }
         else if (name.find("attn_output.weight") != std::string::npos) {
-            if (qs.model.hparams.n_expert == 8) {
-                new_type = GGML_TYPE_Q5_K;
+            if (qs.model.hparams.n_expert >= 8) {
+                new_type = (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
             } else {
                 if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S;
@@ -304,7 +453,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         }
         ++qs.i_attention_wv;
     } else if (name.find("attn_k.weight") != std::string::npos) {
-        if (qs.model.hparams.n_expert == 8) {
+        if (qs.model.hparams.n_expert >= 8) {
             // for the 8-expert model, bumping this to Q8_0 trades just ~128MB
             // TODO: explore better strategies
             new_type = GGML_TYPE_Q8_0;
@@ -322,6 +471,48 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
             new_type = GGML_TYPE_IQ2_S;
         }
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_down_shexp.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        //if (qs.i_ffn_down_shexp < qs.n_ffn_down_shexp/8 || qs.i_ffn_down_shexp > 7*qs.n_ffn_down_shexp/8) {
+        if (use_more_bits(qs.i_ffn_down_shexp, qs.n_ffn_down_shexp)) {
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
+        ++qs.i_ffn_down_shexp;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_gate_shexp.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        //if (qs.i_ffn_gate_shexp < qs.n_ffn_gate_shexp/8 || qs.i_ffn_gate_shexp > 7*qs.n_ffn_gate_shexp/8) {
+        if (use_more_bits(qs.i_ffn_gate_shexp, qs.n_ffn_gate_shexp)) {
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
+        ++qs.i_ffn_gate_shexp;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_up_shexp.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        //if (qs.i_ffn_up_shexp < qs.n_ffn_up_shexp/8 || qs.i_ffn_up_shexp > 7*qs.n_ffn_up_shexp/8) {
+        if (use_more_bits(qs.i_ffn_up_shexp, qs.n_ffn_up_shexp)) {
+            new_type = GGML_TYPE_Q8_0;
+        }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
+        ++qs.i_ffn_up_shexp;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("ffn_down_exps.weight") != std::string::npos) {
+        if (use_more_bits(qs.i_ffn_down_exps, qs.n_ffn_down_exps)) {
+            if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0) && qs.has_imatrix) {
+                // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
+                // We only do it when an imatrix is provided because a) we want to make sure that one can always get the
+                // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
+                new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
+            } else {
+                new_type = get_exps_type_high_bpw_bump(ftype, new_type, qs.has_imatrix);
+            }
+        }
+        ++qs.i_ffn_down_exps;
     } else if (name.find("ffn_down") != std::string::npos) {
         auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
         int i_layer = info.first, n_layer = info.second;
@@ -369,7 +560,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         ++qs.i_ffn_down;
     } else if (name.find("attn_output.weight") != std::string::npos) {
         if (arch != LLM_ARCH_FALCON) {
-            if (qs.model.hparams.n_expert == 8) {
+            if (qs.model.hparams.n_expert >= 8) {
                 if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S  ||
@@ -409,6 +600,24 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
             new_type = GGML_TYPE_IQ3_XXS;
         }
         ++qs.i_ffn_up;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("attn_kv_a_mqa.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q8_0;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("attn_k_b.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("attn_v_b.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
+    } else if (qs.model.hparams.n_expert >= 8 &&name.find("attn_q_b.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q4_K;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
+    } else if (qs.model.hparams.n_expert >= 8 && name.find("attn_q_a.weight") != std::string::npos) {
+        new_type = GGML_TYPE_Q5_K;
+        if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q6_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q8_0;
     }
 
     //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
@@ -700,6 +909,18 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             ++qs.n_attention_wv;
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
+        } else if (name.find("ffn_gate_exps.weight") != std::string::npos) {
+            ++qs.n_ffn_gate_exps;
+        } else if (name.find("ffn_gate_shexp.weight") != std::string::npos) {
+            ++qs.n_ffn_gate_shexp;
+        } else if (name.find("ffn_down_exps.weight") != std::string::npos) {
+            ++qs.n_ffn_down_exps;
+        } else if (name.find("ffn_down_shexp.weight") != std::string::npos) {
+            ++qs.n_ffn_down_shexp;
+        } else if (name.find("ffn_up_exps.weight") != std::string::npos) {
+            ++qs.n_ffn_up_exps;
+        } else if (name.find("ffn_up_shexp.weight") != std::string::npos) {
+            ++qs.n_ffn_up_shexp;
         }
     }