diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 08126dedf7..410305d138 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -446,6 +446,13 @@ class LlamaCppBackend:
         Parses only the KV pairs we need (~30ms even for multi-GB files).
         For split GGUFs, metadata is always in shard 1.
         """
+        # Reset metadata from any previously loaded model so stale flags
+        # (eg _supports_reasoning) do not carry over when switching models.
+        self._context_length = None
+        self._chat_template = None
+        self._supports_reasoning = False
+        self._supports_tools = False
+
         try:
             WANTED = {"general.architecture", "tokenizer.chat_template"}
             arch = None
diff --git a/studio/backend/utils/datasets/llm_assist.py b/studio/backend/utils/datasets/llm_assist.py
index 5012c8ba07..1dc0c495e5 100644
--- a/studio/backend/utils/datasets/llm_assist.py
+++ b/studio/backend/utils/datasets/llm_assist.py
@@ -26,7 +26,7 @@ from loggers import get_logger
 
 logger = get_logger(__name__)
 
-DEFAULT_HELPER_MODEL_REPO = "unsloth/Qwen3-4B-Instruct-2507-GGUF"
+DEFAULT_HELPER_MODEL_REPO = "unsloth/Qwen3.5-4B-GGUF"
 DEFAULT_HELPER_MODEL_VARIANT = "UD-Q4_K_XL"
 
 README_MAX_CHARS = 1500