diff --git a/cli-config.yaml.example b/cli-config.yaml.example
index 657423679..b1d7a67a9 100644
--- a/cli-config.yaml.example
+++ b/cli-config.yaml.example
@@ -348,7 +348,7 @@ compression:
 # Other providers pick a sensible default automatically.
 #
 # auxiliary:
-#   # Image analysis: vision_analyze tool + browser screenshots
+#   # Image analysis: vision_analyze tool
 #   vision:
 #     provider: "auto"
 #     model: ""              # e.g. "google/gemini-2.5-flash", "openai/gpt-4o"
@@ -356,6 +356,13 @@ compression:
 #     download_timeout: 30   # Image HTTP download timeout (seconds)
 #                            # Increase for slow connections or self-hosted image servers
 #
+#   # Browser screenshot analysis (browser_vision tool)
+#   # Defaults to Gemma 4 27B — natively multimodal, same model family as the main
+#   # text model, which avoids model-switching overhead and improves context continuity.
+#   # Override with any vision-capable model.  Set to "" to fall back to auto-detection.
+#   browser_vision:
+#     model: "google/gemma-4-27b-it"  # default; override e.g. "google/gemini-2.5-flash"
+#
 #   # Web page scraping / summarization + browser page text extraction
 #   web_extract:
 #     provider: "auto"
diff --git a/tests/tools/test_browser_vision_model.py b/tests/tools/test_browser_vision_model.py
new file mode 100644
index 000000000..e565da89e
--- /dev/null
+++ b/tests/tools/test_browser_vision_model.py
@@ -0,0 +1,98 @@
+"""Tests for browser_tool._get_vision_model() — Gemma 4 default (Issue #816)."""
+
+import importlib
+import sys
+from unittest.mock import patch, MagicMock
+
+import pytest
+
+
+def _reload_and_get(monkeypatch, env: dict, cfg: dict | None = None):
+    """Reload browser_tool with patched env vars and optional config, return _get_vision_model."""
+    # Patch environment
+    for k, v in env.items():
+        monkeypatch.setenv(k, v)
+
+    # Patch load_config if a config dict is provided
+    if cfg is not None:
+        mock_load = MagicMock(return_value=cfg)
+        monkeypatch.setattr("hermes_cli.config.load_config", mock_load, raising=False)
+
+    # We import the function directly rather than reloading the heavy module
+    # to keep tests fast.  Import browser_tool once and call the function with
+    # patched globals each time.
+    import tools.browser_tool as bt
+    return bt._get_vision_model
+
+
+class TestGetVisionModelDefault:
+    def test_default_is_gemma4(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.delenv("AUXILIARY_VISION_MODEL", raising=False)
+        import tools.browser_tool as bt
+        model = bt._get_vision_model()
+        assert model == "google/gemma-4-27b-it"
+
+    def test_default_constant(self):
+        import tools.browser_tool as bt
+        assert bt._BROWSER_VISION_DEFAULT_MODEL == "google/gemma-4-27b-it"
+
+
+class TestGetVisionModelEnvOverrides:
+    def test_browser_vision_model_env_takes_priority(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_VISION_MODEL", "openai/gpt-4o")
+        monkeypatch.setenv("AUXILIARY_VISION_MODEL", "google/gemini-3-flash-preview")
+        import tools.browser_tool as bt
+        assert bt._get_vision_model() == "openai/gpt-4o"
+
+    def test_auxiliary_vision_model_fallback(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.setenv("AUXILIARY_VISION_MODEL", "google/gemini-3-flash-preview")
+        import tools.browser_tool as bt
+        assert bt._get_vision_model() == "google/gemini-3-flash-preview"
+
+    def test_browser_vision_model_empty_falls_through(self, monkeypatch):
+        """Empty BROWSER_VISION_MODEL should fall through to next step."""
+        monkeypatch.setenv("BROWSER_VISION_MODEL", "")
+        monkeypatch.delenv("AUXILIARY_VISION_MODEL", raising=False)
+        import tools.browser_tool as bt
+        # Should reach the default
+        assert bt._get_vision_model() == "google/gemma-4-27b-it"
+
+    def test_auxiliary_vision_model_empty_falls_through(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.setenv("AUXILIARY_VISION_MODEL", "")
+        import tools.browser_tool as bt
+        assert bt._get_vision_model() == "google/gemma-4-27b-it"
+
+
+class TestGetVisionModelConfig:
+    def test_config_overrides_default(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.delenv("AUXILIARY_VISION_MODEL", raising=False)
+        cfg = {"auxiliary": {"browser_vision": {"model": "anthropic/claude-3-5-haiku"}}}
+        with patch("hermes_cli.config.load_config", return_value=cfg):
+            import tools.browser_tool as bt
+            assert bt._get_vision_model() == "anthropic/claude-3-5-haiku"
+
+    def test_config_empty_string_falls_through_to_default(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.delenv("AUXILIARY_VISION_MODEL", raising=False)
+        cfg = {"auxiliary": {"browser_vision": {"model": ""}}}
+        with patch("hermes_cli.config.load_config", return_value=cfg):
+            import tools.browser_tool as bt
+            assert bt._get_vision_model() == "google/gemma-4-27b-it"
+
+    def test_config_load_error_falls_through_to_default(self, monkeypatch):
+        monkeypatch.delenv("BROWSER_VISION_MODEL", raising=False)
+        monkeypatch.delenv("AUXILIARY_VISION_MODEL", raising=False)
+        with patch("hermes_cli.config.load_config", side_effect=Exception("config error")):
+            import tools.browser_tool as bt
+            assert bt._get_vision_model() == "google/gemma-4-27b-it"
+
+    def test_env_beats_config(self, monkeypatch):
+        monkeypatch.setenv("BROWSER_VISION_MODEL", "openai/gpt-4o")
+        cfg = {"auxiliary": {"browser_vision": {"model": "anthropic/claude-3-5-haiku"}}}
+        with patch("hermes_cli.config.load_config", return_value=cfg):
+            import tools.browser_tool as bt
+            assert bt._get_vision_model() == "openai/gpt-4o"
diff --git a/tools/browser_tool.py b/tools/browser_tool.py
index 03be84e02..9adc20c7e 100644
--- a/tools/browser_tool.py
+++ b/tools/browser_tool.py
@@ -200,9 +200,50 @@ def _get_command_timeout() -> int:
     return result
 
 
+# Default vision model for browser screenshot analysis.
+# Gemma 4 is natively multimodal so it can analyze screenshots using the same
+# model already loaded for text tasks, reducing cold-start latency.
+_BROWSER_VISION_DEFAULT_MODEL = "google/gemma-4-27b-it"
+
+
 def _get_vision_model() -> Optional[str]:
-    """Model for browser_vision (screenshot analysis — multimodal)."""
-    return os.getenv("AUXILIARY_VISION_MODEL", "").strip() or None
+    """Model for browser_vision (screenshot analysis — multimodal).
+
+    Resolution order (first non-empty value wins):
+    1. ``BROWSER_VISION_MODEL`` env var  — browser-specific override
+    2. ``auxiliary.browser_vision.model`` in config.yaml
+    3. ``AUXILIARY_VISION_MODEL`` env var — shared vision override
+    4. ``_BROWSER_VISION_DEFAULT_MODEL``  — Gemma 4 27B (default)
+
+    Set ``BROWSER_VISION_MODEL`` or ``auxiliary.browser_vision.model`` to an
+    empty string to force the auxiliary router's auto-detection (no default).
+    """
+    # 1. Browser-specific env var
+    env_browser = os.getenv("BROWSER_VISION_MODEL", "").strip()
+    if env_browser:
+        return env_browser
+
+    # 2. Config file: auxiliary.browser_vision.model
+    try:
+        from hermes_cli.config import load_config
+        _cfg = load_config()
+        cfg_model = (
+            _cfg.get("auxiliary", {})
+            .get("browser_vision", {})
+            .get("model", "")
+        )
+        if cfg_model and str(cfg_model).strip():
+            return str(cfg_model).strip()
+    except Exception:
+        pass
+
+    # 3. Shared vision env var (backward-compat)
+    env_shared = os.getenv("AUXILIARY_VISION_MODEL", "").strip()
+    if env_shared:
+        return env_shared
+
+    # 4. Default: Gemma 4 27B
+    return _BROWSER_VISION_DEFAULT_MODEL
 
 
 def _get_extraction_model() -> Optional[str]:
@@ -1893,20 +1934,23 @@ def browser_get_images(task_id: Optional[str] = None) -> str:
 def browser_vision(question: str, annotate: bool = False, task_id: Optional[str] = None) -> str:
     """
     Take a screenshot of the current page and analyze it with vision AI.
-    
-    This tool captures what's visually displayed in the browser and sends it
-    to Gemini for analysis. Useful for understanding visual content that the
-    text-based snapshot may not capture (CAPTCHAs, verification challenges,
-    images, complex layouts, etc.).
-    
+
+    Uses Gemma 4 27B by default (natively multimodal — same model family as the
+    main text model, lower cold-start latency than switching to a separate vision
+    model).  Override via ``BROWSER_VISION_MODEL`` env var or
+    ``auxiliary.browser_vision.model`` in config.yaml.
+
+    Useful for understanding visual content that the text-based snapshot may not
+    capture (CAPTCHAs, verification challenges, images, complex layouts, etc.).
+
     The screenshot is saved persistently and its file path is returned alongside
     the analysis, so it can be shared with users via MEDIA:<path> in the response.
-    
+
     Args:
         question: What you want to know about the page visually
         annotate: If True, overlay numbered [N] labels on interactive elements
         task_id: Task identifier for session isolation
-        
+
     Returns:
         JSON string with vision analysis results and screenshot_path
     """