diff --git a/aiserver.py b/aiserver.py
index fe5e5aa2..3e6b5a3e 100644
--- a/aiserver.py
+++ b/aiserver.py
@@ -1916,9 +1916,6 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
     if koboldai_vars.model == "ReadOnly":
         koboldai_vars.noai = True
 
-    loadmodelsettings()
-    loadsettings()
-
     # TODO: InferKit
     if koboldai_vars.model == "ReadOnly" or koboldai_vars.noai:
         pass
@@ -1984,6 +1981,9 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
     if model:
         tokenizer = model.tokenizer
 
+    loadmodelsettings()
+    loadsettings()
+
     lua_startup()
     # Load scripts
     load_lua_scripts()
diff --git a/modeling/inference_model.py b/modeling/inference_model.py
index 0be06d17..b253c5bf 100644
--- a/modeling/inference_model.py
+++ b/modeling/inference_model.py
@@ -197,12 +197,6 @@ class InferenceModel:
         Returns:
             AutoTokenizer: Tokenizer deemed fit for the location string. May be a fallback tokenizer.
         """
-        if utils.koboldai_vars.model_type == "xglm":
-            # Default to </s> newline mode if using XGLM
-            utils.koboldai_vars.newlinemode = "s"
-        elif utils.koboldai_vars.model_type in ["opt", "bloom"]:
-            # Handle </s> but don't convert newlines if using Fairseq models that have newlines trained in them
-            utils.koboldai_vars.newlinemode = "ns"
 
         std_kwargs = {"revision": utils.koboldai_vars.revision, "cache_dir": "cache"}
 
diff --git a/modeling/inference_models/hf.py b/modeling/inference_models/hf.py
index eac5284f..013590ef 100644
--- a/modeling/inference_models/hf.py
+++ b/modeling/inference_models/hf.py
@@ -18,6 +18,23 @@ class HFInferenceModel(InferenceModel):
         self.tokenizer = None
 
     def _post_load(self) -> None:
+        # These are model specific tokenizer overrides if a model has bad defaults
+        if utils.koboldai_vars.model_type == "llama":
+            self.tokenizer.decode_with_prefix_space = True
+            self.tokenizer.add_bos_token = False
+        elif utils.koboldai_vars.model_type == "opt":
+            self.tokenizer._koboldai_header = self.tokenizer.encode("")
+            self.tokenizer.add_bos_token = False
+            self.tokenizer.add_prefix_space = False
+
+        # Change newline behavior to match model quirks
+        if utils.koboldai_vars.model_type == "xglm":
+            # Default to </s> newline mode if using XGLM
+            utils.koboldai_vars.newlinemode = "s"
+        elif utils.koboldai_vars.model_type in ["opt", "bloom"]:
+            # Handle </s> but don't convert newlines if using Fairseq models that have newlines trained in them
+            utils.koboldai_vars.newlinemode = "ns"
+
         # Clean up tokens that cause issues
         if (
             utils.koboldai_vars.badwordsids == koboldai_settings.badwordsids_default
diff --git a/modeling/inference_models/hf_torch.py b/modeling/inference_models/hf_torch.py
index 1997e7fe..49cdfc0f 100644
--- a/modeling/inference_models/hf_torch.py
+++ b/modeling/inference_models/hf_torch.py
@@ -132,15 +132,6 @@ class HFTorchInferenceModel(HFInferenceModel):
         if not utils.koboldai_vars.model_type:
             utils.koboldai_vars.model_type = m_self.get_model_type()
 
-        # These are model specific overrides if a model has bad defaults
-        if utils.koboldai_vars.model_type == "llama":
-            m_self.tokenizer.decode_with_prefix_space = True
-            m_self.tokenizer.add_bos_token = False
-        elif utils.koboldai_vars.model_type == "opt":
-            m_self.tokenizer._koboldai_header = m_self.tokenizer.encode("")
-            m_self.tokenizer.add_bos_token = False
-            m_self.tokenizer.add_prefix_space = False
-
         # Patch stopping_criteria
         class PTHStopper(StoppingCriteria):
             def __call__(