through divine intervention it was revealed to me a couple of critical bugs and with the help of Claude + an updated model I've finally achieved faster and smarter responses

main.py

@@ -1,20 +1,15 @@
 #!/usr/bin/env python3
 import asyncio
 import json
+import re  # thank you claude
 import ollama
 import websockets
 from websockets.server import serve
 from datetime import datetime
 
-chatmodel = "hf.co/Orenguteng/Llama-3.1-8B-Lexi-Uncensored-V2-GGUF:Llama-3.1-8B-Lexi-Uncensored_V2_Q4.gguf"
+chatmodel = "hf.co/Stetsoncody/Llama-3.2-3B-Instruct-abliterated-GGUF:Q4_K_M"
 embedmodel = 'hf.co/CompendiumLabs/bge-base-en-v1.5-gguf'
 
-# these are test models just ignore
-#chatmodel = 'hf.co/bartowski/Llama-3.2-1B-Instruct-GGUF'
-#chatmodel = 'hf.co/VibeStudio/Nidum-Llama-3.2-3B-Uncensored-GGUF'
-#chatmodel = "faggotchan"
-#chatmodel = "qwen3.5:4b"
-
 NUM_THREADS = 6  # hardcoded since threads.txt is gone, change if needed
 
 def times() -> str:
@@ -160,16 +155,17 @@ async def ollamachat(websocket):
             model=chatmodel,
             messages=messages,
             options={
-                "temperature":    0.8,
-                "top_p":          0.9,
+                "temperature":    0.83,
+                "top_p":          0.9
                 "top_k":          40,
                 "repeat_penalty": 1.1,
-                "num_predict":    4096,
+                "num_predict":    1024,
                 "num_thread":     NUM_THREADS,
             }
         )
 
-        bot_reply = response["message"]["content"].strip()
+        # strip <think> blocks in case the model ignores think=False, thank you claude
+        bot_reply = re.sub(r'<think>.*?</think>', '', response["message"]["content"], flags=re.DOTALL).strip()
 
         print(messages)
         print(f"[{times()}] reply: {bot_reply}\n")