feat: Switch models to use Gemma 4 from unsloth collection

firefox-extension/popup.html

@@ -100,8 +100,8 @@
 
   <label for="model-version">Model Version</label>
   <select id="model-version" style="width: 100%; margin-bottom: 15px; padding: 8px; border: 1px solid #d6d3d1; border-radius: 4px; font-family: inherit;">
-    <option value="2b">Gemma 2B (Fast / Default)</option>
-    <option value="27b-4a">Gemma 27B 4-bit (High Quality)</option>
+    <option value="2b">Gemma 4 2B (Fast / Default)</option>
+    <option value="27b-4a">Gemma 4 27B 4-bit (High Quality)</option>
   </select>
   <p class="help" style="margin-top: 0;">Select which LLM to use for saliency analysis.</p>
 

main.py

@@ -247,8 +247,8 @@ def load_model(model_name: str):
     print(f"Loading {model_name} on {device}...")
     try:
         if model_name == "27b-4a":
-            # Use Gemma 2 27B in 4-bit (requires CUDA)
-            hf_model_id = "google/gemma-2-27b"
+            # Use Gemma 4 27B in 4-bit (requires CUDA)
+            hf_model_id = "unsloth/gemma-4-27b"
             tokenizer = AutoTokenizer.from_pretrained(hf_model_id, token=hf_token)
             
             # BitsAndBytes for 4-bit quantization
@@ -267,7 +267,7 @@ def load_model(model_name: str):
             )
         else:
             # Default to 2b
-            hf_model_id = "google/gemma-2b"
+            hf_model_id = "unsloth/gemma-4-2b"
             tokenizer = AutoTokenizer.from_pretrained(hf_model_id, token=hf_token)
             model = AutoModelForCausalLM.from_pretrained(
                 hf_model_id,