Update app.py

app.py

@@ -1,9 +1,9 @@
 """
 Text-to-Game Generator - Local GGUF Edition
-- Code  : Qwen/Qwen2.5-Coder-7B-Instruct-GGUF Q4_K_M (~4.5GB, CPU via llama-cpp-python)
-- Images: unsloth/Z-Image-Turbo-GGUF Q2_K       (~3.6GB, GPU via ZeroGPU + diffusers)
+- Code  : WithinUsAI/Opus4.7-GODs.Ghost.Codex-4B.GGuF (~2-3GB, CPU via llama-cpp-python)
+- Images: unsloth/Z-Image-Turbo-GGUF Q2_K             (~3.6GB, GPU via ZeroGPU + diffusers)
 
-Total disk usage: ~8GB — fits within HF free Space 16GB storage limit.
+Total disk usage: ~6GB — well within HF free Space 16GB storage limit.
 Space hardware   : set to ZeroGPU (free) in Space settings.
 
 Models are downloaded on first startup and cached — subsequent starts are instant.
@@ -25,22 +25,24 @@ from PIL import Image
 # Model paths — downloaded once at startup, cached in /tmp
 # ---------------------------------------------------------------------------
 
-GGUF_DIR         = Path("/tmp/models")
-CODE_GGUF_PATH   = GGUF_DIR / "qwen2.5-coder-7b-q4_k_m.gguf"
-IMAGE_GGUF_PATH  = GGUF_DIR / "z-image-turbo-Q2_K.gguf"
+GGUF_DIR        = Path("/tmp/models")
+IMAGE_GGUF_PATH = GGUF_DIR / "z-image-turbo-Q2_K.gguf"
 
 GGUF_DIR.mkdir(parents=True, exist_ok=True)
 
-# Download code model GGUF (Q4_K_M ~4.5GB, good quality/speed balance on CPU)
+# Download code model — auto-detect filename from repo
+CODE_REPO      = "WithinUsAI/Opus4.7-GODs.Ghost.Codex-4B.GGuF"
+CODE_FILENAME  = "Opus4.7-Distill-GODsGhost-Codex-4B-Q4_K_M.gguf"
+CODE_GGUF_PATH = GGUF_DIR / CODE_FILENAME
+
 if not CODE_GGUF_PATH.exists():
-    print("Downloading Qwen2.5-Coder-7B Q4_K_M GGUF...")
+    print("Downloading Opus4.7-GODs.Ghost.Codex-4B Q4_K_M GGUF...")
     hf_hub_download(
-        repo_id="Qwen/Qwen2.5-Coder-7B-Instruct-GGUF",
-        filename="qwen2.5-coder-7b-instruct-q4_k_m.gguf",
+        repo_id=CODE_REPO,
+        filename=CODE_FILENAME,
         local_dir=str(GGUF_DIR),
         local_dir_use_symlinks=False,
     )
-    (GGUF_DIR / "qwen2.5-coder-7b-instruct-q4_k_m.gguf").rename(CODE_GGUF_PATH)
     print("Code model downloaded.")
 
 # Download image model GGUF (Q2_K ~3.6GB, smallest that still works)
@@ -58,14 +60,15 @@ if not IMAGE_GGUF_PATH.exists():
 # Load code model (llama-cpp-python, CPU, loaded once at startup)
 # ---------------------------------------------------------------------------
 
-print("Loading Qwen2.5-Coder-7B into llama-cpp...")
+print("Loading Opus4.7-GODs.Ghost.Codex-4B into llama-cpp...")
 from llama_cpp import Llama
 
-code_llm = Llama(
-    model_path=str(CODE_GGUF_PATH),
-    n_ctx=4096,        # context window
-    n_threads=4,       # CPU threads
-    n_gpu_layers=0,    # CPU only
+code_llm = Llama.from_pretrained(
+    repo_id=CODE_REPO,
+    filename=CODE_FILENAME,
+    n_ctx=4096,
+    n_threads=4,
+    n_gpu_layers=0,
     verbose=False,
 )
 print("Code model ready.")
@@ -291,7 +294,7 @@ def generate_game_code(game_type: str, theme: str, temperature: float, max_new_t
     )
 
     try:
-        # Step 1: Generate HTML via Qwen2.5-Coder-7B (llama-cpp, CPU)
+        # Step 1: Generate HTML via Opus4.7-GODs.Ghost.Codex-4B (llama-cpp, CPU)
         response = code_llm.create_chat_completion(
             messages=[
                 {"role": "system", "content": system_msg},
@@ -390,7 +393,7 @@ def build_ui():
 
         gr.Markdown(
             "# Game Generator (Local GGUF)\n"
-            "**Qwen2.5-Coder-7B Q4** generates the game code (CPU). "
+            "**Opus4.7-GODs.Ghost.Codex-4B** generates the game code (CPU). "
             "**Z-Image-Turbo Q2** generates sprites (ZeroGPU).\n"
             "Both models run locally — no API quota, no credits needed.\n\n"
             "> First generation takes a few minutes while models download and load. "
@@ -467,7 +470,7 @@ def build_ui():
 
         gr.Markdown(
             "---\n"
-            "**Code:** Qwen2.5-Coder-7B-Instruct Q4_K_M GGUF (~4.5GB) via llama-cpp on CPU. "
+            "**Code:** Opus4.7-GODs.Ghost.Codex-4B GGUF (~2-3GB) via llama-cpp on CPU. "
             "**Images:** Z-Image-Turbo Q2_K GGUF (~3.6GB) via diffusers on ZeroGPU. "
             "Total disk: ~8GB of the 16GB free limit. No API calls, no quota."
         )