diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 2e4645e81c73..ca03d09a8291 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -7388,6 +7388,8 @@ def set_gguf_parameters(self):
         rope_dim = int(self.hparams["head_dim"] * self.hparams["partial_rotary_factor"])
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon", 1e-5))
+
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch, name, bid):
@@ -7433,6 +7435,15 @@ def modify_tensors(self, data_torch, name, bid):
                 return []
         return [(self.map_tensor_name(name), data_torch)]
 
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
 
 @ModelBase.register("PanguEmbeddedForCausalLM")
 class PanguEmbeddedModel(TextModel):
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 1466283c74e0..a32e1c2806dd 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -2341,6 +2341,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_MIMO2:
             {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.rope_freq_base_train_swa = 10000.0f;
 
diff --git a/src/models/mimo2-iswa.cpp b/src/models/mimo2-iswa.cpp
index 3335fe841df8..edc87cc9f0d3 100644
--- a/src/models/mimo2-iswa.cpp
+++ b/src/models/mimo2-iswa.cpp
@@ -88,17 +88,10 @@ llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_
             cb(cur, "ffn_out", il);
         } else {
             // MoE branch
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
-                    il);
+            cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
+                                model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
+                                model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false,
+                                0.0, LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, il);
             cb(cur, "ffn_moe_out", il);
         }