feat: add Qwen2.5 3B and Qwen3 4B extraction models

- Added Qwen2.5 3B: Qwen/Qwen2.5-3B-Instruct-GGUF
- Added Qwen3 4B: unsloth/Qwen3-4B-GGUF (supports reasoning)
- Set Qwen3 4B as DEFAULT_EXTRACTION_MODEL
- Both configured with temperature 0.1 for greedy extraction

These larger models should handle Chinese transcript extraction better
than Granite 4.0 Tiny which only achieved 3.4% success rate.

app.py

@@ -863,9 +863,37 @@ EXTRACTION_MODELS = {
             "repeat_penalty": 1.0,
         },
     },
+    "qwen2.5_3b": {
+        "name": "Qwen2.5 3B (128K Context)",
+        "repo_id": "Qwen/Qwen2.5-3B-Instruct-GGUF",
+        "filename": "*Q4_K_M.gguf",
+        "max_context": 131072,
+        "supports_reasoning": False,
+        "supports_toggle": False,
+        "inference_settings": {
+            "temperature": 0.1,
+            "top_p": 0.9,
+            "top_k": 20,
+            "repeat_penalty": 1.0,
+        },
+    },
+    "qwen3_4b": {
+        "name": "Qwen3 4B (32K Context)",
+        "repo_id": "unsloth/Qwen3-4B-GGUF",
+        "filename": "*Q4_K_M.gguf",
+        "max_context": 32768,
+        "supports_reasoning": True,
+        "supports_toggle": True,
+        "inference_settings": {
+            "temperature": 0.1,
+            "top_p": 0.9,
+            "top_k": 20,
+            "repeat_penalty": 1.0,
+        },
+    },
 }
 
-DEFAULT_EXTRACTION_MODEL = "granite4_tiny_q3"
+DEFAULT_EXTRACTION_MODEL = "qwen3_4b"
 
 
 # ===== ADVANCED MODE: SYNTHESIS MODELS REGISTRY (16 models, 1B-30B) =====