diff --git a/conversion/__init__.py b/conversion/__init__.py
index 3976e03..4c7d5cc 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -213,6 +213,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Ministral3Model": "mistral3",
     "Mistral3ForConditionalGeneration": "mistral3",
     "Mistral3Model": "mistral3",
+    "Mistral3ForSequenceClassification": "mistral3",
+    "Ministral3ForSequenceClassification": "mistral3",
     "MistralLarge4ForConditionalGeneration": "mistral4",
     "VoxtralRealtimeForConditionalGeneration": "mistral3",
     "MistralForCausalLM": "llama",
diff --git a/conversion/mistral3.py b/conversion/mistral3.py
index 6c771b2..e77bef4 100644
--- a/conversion/mistral3.py
+++ b/conversion/mistral3.py
@@ -16,6 +16,8 @@ from .llama import LlamaModel
     "Mistral3Model",
     "Ministral3ForCausalLM",
     "Ministral3Model",
+    "Mistral3ForSequenceClassification",  # LOCAL PATCH
+    "Ministral3ForSequenceClassification",  # LOCAL PATCH
 )
 @ModelBase.example("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM", "nvidia/Nemotron-3-Embed-1B-BF16", "nvidia/Nemotron-3.5-Embed-1B-BF16")
 class Mistral3Model(TextModel):
@@ -40,13 +42,28 @@ class Mistral3Model(TextModel):
             self.gguf_writer.add_add_eos_token(False)
 
         def get_vocab_base_pre(self, tokenizer) -> str:
-            if self.origin_hf_arch == "Mistral3Model":
+            if self.origin_hf_arch in ("Mistral3Model", "Mistral3ForSequenceClassification", "Ministral3ForSequenceClassification"):
                 return "nemotron-3.5-embed"
             return super().get_vocab_base_pre(tokenizer)
 
         def set_gguf_parameters(self):
             super().set_gguf_parameters()
             self._try_set_pooling_type()
+            if self.is_rerank:  # LOCAL PATCH: Nemotron rerank -> native RANK pooling (mean + cls.output)
+                self.gguf_writer.add_pooling_type(gguf.PoolingType.RANK)
+                self.gguf_writer.add_classifier_output_labels(["score"])
+                self.gguf_writer.add_chat_template([{"name": "rerank", "template": "query: {query}\n\npassage: {document}"}])
+
+        @property
+        def is_rerank(self) -> bool:
+            return "sequenceclassification" in (self.origin_hf_arch or "").lower()
+
+        def modify_tensors(self, data_torch, name, bid):  # LOCAL PATCH: rerank score head -> cls.output
+            if name.endswith("score.weight"):
+                if self.is_rerank:
+                    yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.CLS_OUT] + ".weight", data_torch)
+                return
+            yield from super().modify_tensors(data_torch, name, bid)
 
     class Mistral4Model(DeepseekV2Model):
         model_arch = gguf.MODEL_ARCH.MISTRAL4
@@ -64,14 +81,14 @@ class Mistral3Model(TextModel):
     def __init__(self, dir_model, *args, **kwargs):
         hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
         hf_arch = get_model_architecture(hparams, self.model_type)
-        if hf_arch in ("Ministral3Model", "Mistral3Model"):
+        if hf_arch in ("Ministral3Model", "Mistral3Model", "Mistral3ForSequenceClassification", "Ministral3ForSequenceClassification"):
             text_config = hparams.get("text_config", hparams)
             if text_config.get("is_causal") is not False:
                 raise NotImplementedError(f"Expected a Nemotron Embed model for {hf_arch}, but is_causal is not False")
         super().__init__(dir_model, *args, **kwargs)
         if self.hparams.get("model_type") == "mistral4":
             self.impl = Mistral3Model.Mistral4Model(dir_model, *args, **kwargs)
-        elif self.hf_arch in ("Ministral3Model", "Mistral3Model"):
+        elif self.hf_arch in ("Ministral3Model", "Mistral3Model", "Mistral3ForSequenceClassification", "Ministral3ForSequenceClassification"):
             self.impl = Mistral3Model.Mistral3EmbedNemotronModel(dir_model, *args, **kwargs)
         else:
             self.impl = Mistral3Model.Ministral3Model(dir_model, *args, **kwargs)
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 1d67dc8..238b26a 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -5820,6 +5820,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     ],
     MODEL_ARCH.MISTRAL3: [
         MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.CLS_OUT,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,
         MODEL_TENSOR.ROPE_FREQS,
diff --git a/src/models/mistral3.cpp b/src/models/mistral3.cpp
index af058d5..bbdad09 100644
--- a/src/models/mistral3.cpp
+++ b/src/models/mistral3.cpp
@@ -20,6 +20,11 @@ void llama_model_mistral3::load_arch_hparams(llama_model_loader & ml) {
         }
     }
 
+    // rank pooling (Nemotron rerank): mean pooling before cls_out
+    if (hparams.pooling_type_cls == LLAMA_POOLING_TYPE_UNSPECIFIED) {
+        hparams.pooling_type_cls = LLAMA_POOLING_TYPE_MEAN;
+    }
+
     switch (hparams.n_layer()) {
         case 26: type = LLM_TYPE_3B; break;
         case 34: type = LLM_TYPE_8B; break;
@@ -38,6 +43,9 @@ void llama_model_mistral3::load_arch_tensors(llama_model_loader & ml) {
     output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
 
     // if output is NULL, init from the input tok embed
+    // optional classifier head (rerank)
+    cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+
     if (output == NULL) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
     }