Bump default Gemini model to gemini-3.5-flash

.github/workflows/nightly-examples.yml

@@ -17,7 +17,7 @@ env:
   NOTTE_API_KEY: ${{ secrets.NOTTE_API_KEY }}
   GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
   # Use the Vertex credentials configured below for the local-agent example.
-  NOTTE_EXAMPLE_MODEL: "vertex_ai/gemini-2.5-flash"
+  NOTTE_EXAMPLE_MODEL: "vertex_ai/gemini-3.5-flash"
   GITHUB_COM_EMAIL: ${{ secrets.BOT_GITHUB_COM_EMAIL }}
   GITHUB_COM_PASSWORD: ${{ secrets.BOT_GITHUB_COM_PASSWORD }}
   GITHUB_COM_MFA_SECRET: ${{ secrets.BOT_GITHUB_COM_MFA_SECRET }}

README.md

@@ -54,7 +54,7 @@ from dotenv import load_dotenv
 load_dotenv()
 
 with notte.Session(headless=False) as session:
-    model = os.getenv("NOTTE_EXAMPLE_MODEL", "gemini/gemini-2.5-flash")
+    model = os.getenv("NOTTE_EXAMPLE_MODEL", "gemini/gemini-3.5-flash")
     agent = notte.Agent(session=session, reasoning_model=model, max_steps=10)
     response = agent.run(task="Find three cat memes on Google Images and describe them")
 ```
@@ -70,7 +70,7 @@ import os
 client = NotteClient(api_key=os.getenv("NOTTE_API_KEY"))
 
 with client.Session(open_viewer=True) as session:
-    agent = client.Agent(session=session, reasoning_model='gemini/gemini-2.5-flash', max_steps=30)
+    agent = client.Agent(session=session, reasoning_model='gemini/gemini-3.5-flash', max_steps=30)
     response = agent.run(task="doom scroll cat memes on google images")
 ```
 
@@ -108,7 +108,7 @@ class TopPosts(BaseModel):
 
 client = NotteClient()
 with client.Session(open_viewer=True, browser_type="chrome") as session:
-    agent = client.Agent(session=session, reasoning_model='gemini/gemini-2.5-flash', max_steps=15)
+    agent = client.Agent(session=session, reasoning_model='gemini/gemini-3.5-flash', max_steps=15)
     response = agent.run(
         task="Go to Hacker News (news.ycombinator.com) and extract the top 5 posts with their titles, URLs, points, authors, and comment counts.",
         response_format=TopPosts,

docs/src/snippets/integrations/mcp/google_adk.mdx

@@ -18,7 +18,7 @@ toolset = McpToolset(
 
 agent = LlmAgent(
     name="notte_browser_agent",
-    model="gemini-2.5-flash",
+    model="gemini-3.5-flash",
     instruction="You browse the web using the Notte tools.",
     tools=[toolset],
 )

docs/src/testers/integrations/mcp/google_adk.py

@@ -16,7 +16,7 @@
 
 agent = LlmAgent(
     name="notte_browser_agent",
-    model="gemini-2.5-flash",
+    model="gemini-3.5-flash",
     instruction="You browse the web using the Notte tools.",
     tools=[toolset],
 )

node-sdk/test/agent.integration.test.ts

@@ -59,7 +59,7 @@ describe('Agent Integration Tests', () => {
       await client.Session({ proxies: false, headless: true, idle_timeout_minutes: 2 }).use(async (session) => {
         const agent = client.Agent({
           session,
-          reasoning_model: 'gemini/gemini-2.5-flash',
+          reasoning_model: 'gemini/gemini-3.5-flash',
           max_steps: 3
         });
 

packages/notte-core/src/notte_core/common/config.py

@@ -132,8 +132,8 @@ def has_apikey_in_env(self) -> bool:
 
 class LlmModel(StrEnum):
     openai = "openai/gpt-4o"
-    gemini = "gemini/gemini-2.5-flash"
-    gemini_vertex = "vertex_ai/gemini-2.5-flash"
+    gemini = "gemini/gemini-3.5-flash"
+    gemini_vertex = "vertex_ai/gemini-3.5-flash"
     gemma = "openrouter/google/gemma-3-27b-it"
     cerebras = "cerebras/gpt-oss-120b"
     groq = "groq/gpt-oss-120b"

packages/notte-llm/src/notte_llm/engine.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 import re
 from collections.abc import Iterable
 from dataclasses import dataclass
@@ -166,6 +167,18 @@ def is_gemini_model(model: str) -> bool:
     return "gemini" in model_lower or "vertex_ai" in model_lower
 
 
+def get_vertex_location(model: str) -> str | None:
+    """Vertex AI location to use for a Gemini model, or None for other models.
+
+    litellm defaults to us-central1, but recent Gemini models (e.g. gemini-3.5-flash) are only
+    served from the global endpoint and a few regions. Default to global unless a location is
+    already configured.
+    """
+    if not model.lower().startswith("vertex_ai/gemini"):
+        return None
+    return litellm.vertex_location or os.getenv("VERTEXAI_LOCATION") or os.getenv("VERTEX_LOCATION") or "global"
+
+
 def is_anthropic_model(model: str) -> bool:
     """Check if the model is an Anthropic model."""
     model_lower = model.lower()
@@ -558,6 +571,8 @@ async def completion(
         model = self._get_model(model)
         # Apply model-specific temperature overrides
         temperature = LlmModel.get_temperature(model, temperature)
+        vertex_location = get_vertex_location(model)
+        vertex_kwargs: dict[str, Any] = {"vertex_location": vertex_location} if vertex_location is not None else {}
         try:
             response = await litellm.acompletion(  # pyright: ignore [reportUnknownMemberType]
                 model,
@@ -572,6 +587,7 @@ async def completion(
                 # indefinitely. Without this, httpx has no read timeout and silent server
                 # stalls hang the whole agent run.
                 timeout=60,
+                **vertex_kwargs,
             )
             # Cast to ModelResponse since we know it's not streaming in this case
             return cast(ModelResponse, response)

tests/config/test_default_config.py

@@ -1,5 +1,5 @@
 import pytest
-from notte_core.common.config import NotteConfig, config
+from notte_core.common.config import LlmModel, NotteConfig, config
 from pydantic import ValidationError
 
 
@@ -25,6 +25,16 @@ def test_default_config():
     assert config.evaluate_js_max_result_bytes == 16 * 1024 * 1024
 
 
+def test_default_llm_model_uses_vertex_with_google_credentials(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("GOOGLE_APPLICATION_CREDENTIALS", "credentials.json")
+    assert LlmModel.default() == "vertex_ai/gemini-3.5-flash"
+
+
+def test_default_llm_model_uses_gemini_without_google_credentials(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("GOOGLE_APPLICATION_CREDENTIALS", raising=False)
+    assert LlmModel.default() == "gemini/gemini-3.5-flash"
+
+
 def test_default_is_headless():
     assert config.headless, "headless should be true by default for tests"
 

tests/config/test_openrouter_provider.py

@@ -55,7 +55,7 @@ def test_anthropic_model_returns_none(self) -> None:
         assert LlmModel.get_openrouter_provider("anthropic/claude-sonnet-4-5-20250929") is None
 
     def test_gemini_model_returns_none(self) -> None:
-        assert LlmModel.get_openrouter_provider("gemini/gemini-2.5-flash") is None
+        assert LlmModel.get_openrouter_provider("gemini/gemini-3.5-flash") is None
 
 
 class TestGetOpenrouterModel:
@@ -82,12 +82,12 @@ def test_claude_sonnet_conversion(self) -> None:
         assert result == "openrouter/anthropic/claude-sonnet-4-5"
 
     def test_vertex_ai_conversion(self) -> None:
-        result = LlmModel.get_openrouter_model("vertex_ai/gemini-2.5-flash")
-        assert result == "openrouter/google/gemini-2.5-flash"
+        result = LlmModel.get_openrouter_model("vertex_ai/gemini-3.5-flash")
+        assert result == "openrouter/google/gemini-3.5-flash"
 
     def test_gemini_prefix_conversion(self) -> None:
-        result = LlmModel.get_openrouter_model("gemini/gemini-2.5-flash")
-        assert result == "openrouter/google/gemini-2.5-flash"
+        result = LlmModel.get_openrouter_model("gemini/gemini-3.5-flash")
+        assert result == "openrouter/google/gemini-3.5-flash"
 
     def test_kimi_conversion(self) -> None:
         result = LlmModel.get_openrouter_model("moonshot/kimi-k2.5")

tests/integration/sdk/test_agent.py

@@ -33,7 +33,7 @@ def test_agent_gemini_form_fill_no_null_fields():
     _ = load_dotenv()
     client = NotteClient()
     with client.Session(proxies=False) as session:
-        agent = client.Agent(session=session, max_steps=3, reasoning_model="vertex_ai/gemini-2.5-flash")
+        agent = client.Agent(session=session, max_steps=3, reasoning_model="vertex_ai/gemini-3.5-flash")
         response = agent.run(
             task="Ignore the web page. Simply return a form fill action with email='lucas@notte.cc' and password='123456'. Stop immediately after this",
             url="https://github.com/login",
@@ -49,7 +49,7 @@ def test_local_agent_gemini_form_fill_no_null_fields():
     """Local agent: Gemini should only fill requested fields, not all fields with null."""
     _ = load_dotenv()
     with notte.Session(headless=True) as session:
-        agent = notte.Agent(session=session, max_steps=3, reasoning_model="vertex_ai/gemini-2.5-flash")
+        agent = notte.Agent(session=session, max_steps=3, reasoning_model="vertex_ai/gemini-3.5-flash")
         response = agent.run(
             task="Ignore the web page. Simply return a form fill action with email='lucas@notte.cc' and password='123456'. Stop immediately after this",
             url="https://console.notte.cc/login",
@@ -65,7 +65,7 @@ def test_start_agent_with_gemini_reasoning():
     _ = load_dotenv()
     notte = NotteClient()
     with notte.Session(proxies=False) as session:
-        agent = notte.Agent(session=session, reasoning_model="gemini/gemini-2.5-flash", max_steps=3)
+        agent = notte.Agent(session=session, reasoning_model="gemini/gemini-3.5-flash", max_steps=3)
         _ = agent.run(task="Go notte.cc and describe the page")
     resp = agent.status()
     assert resp.status == "closed"

tests/llms/test_engine.py

@@ -3,7 +3,13 @@
 import pytest
 from litellm import Message
 from notte_core.errors.base import ErrorConfig
-from notte_llm.engine import LLMEngine, StructuredContent, fix_schema_for_gemini, fix_schema_for_openai
+from notte_llm.engine import (
+    LLMEngine,
+    StructuredContent,
+    fix_schema_for_gemini,
+    fix_schema_for_openai,
+    get_vertex_location,
+)
 
 
 @pytest.fixture
@@ -43,6 +49,60 @@ async def test_completion_error(llm_engine: LLMEngine) -> None:
             assert "API Error" in str(exc_info.value)
 
 
+class TestGetVertexLocation:
+    @pytest.fixture(autouse=True)
+    def _no_configured_location(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.setattr("litellm.vertex_location", None)
+        monkeypatch.delenv("VERTEXAI_LOCATION", raising=False)
+        monkeypatch.delenv("VERTEX_LOCATION", raising=False)
+
+    def test_vertex_gemini_defaults_to_global(self) -> None:
+        assert get_vertex_location("vertex_ai/gemini-3.5-flash") == "global"
+
+    def test_env_location_is_respected(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.setenv("VERTEXAI_LOCATION", "europe-west2")
+        assert get_vertex_location("vertex_ai/gemini-3.5-flash") == "europe-west2"
+
+    def test_litellm_location_is_respected(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        monkeypatch.setattr("litellm.vertex_location", "us")
+        assert get_vertex_location("vertex_ai/gemini-3.5-flash") == "us"
+
+    @pytest.mark.parametrize(
+        "model",
+        [
+            "gemini/gemini-3.5-flash",
+            "openrouter/google/gemini-3.5-flash",
+            "vertex_ai/claude-sonnet-4-5",
+            "vertex_ai/mistral-large",
+            "vertex_ai/meta/llama-3.3-70b-instruct-maas",
+            "openai/gpt-4o",
+        ],
+    )
+    def test_other_models_have_no_location(self, model: str) -> None:
+        assert get_vertex_location(model) is None
+
+
+@pytest.mark.asyncio
+async def test_completion_passes_global_location_for_vertex_gemini(
+    llm_engine: LLMEngine, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.setattr("litellm.vertex_location", None)
+    monkeypatch.delenv("VERTEXAI_LOCATION", raising=False)
+    monkeypatch.delenv("VERTEX_LOCATION", raising=False)
+    with patch("litellm.acompletion", return_value=Mock()) as acompletion:
+        await llm_engine.completion(
+            messages=[Message(role="user", content="Hello")], model="vertex_ai/gemini-3.5-flash"
+        )
+    assert acompletion.call_args.kwargs["vertex_location"] == "global"
+
+
+@pytest.mark.asyncio
+async def test_completion_omits_location_for_other_models(llm_engine: LLMEngine) -> None:
+    with patch("litellm.acompletion", return_value=Mock()) as acompletion:
+        await llm_engine.completion(messages=[Message(role="user", content="Hello")], model="gpt-3.5-turbo")
+    assert "vertex_location" not in acompletion.call_args.kwargs
+
+
 class TestStructuredContent:
     def test_extract_with_outer_tag(self):
         structure = StructuredContent(outer_tag="response")

tests/llms/test_openrouter_models.py

@@ -18,7 +18,7 @@
 # Update this list as new models become available
 OPENROUTER_MODELS = [
     "google/gemini-3-flash-preview",
-    "google/gemini-2.5-flash",
+    "google/gemini-3.5-flash",
     "anthropic/claude-opus-4.6",
     "anthropic/claude-sonnet-4.6",
     "anthropic/claude-haiku-4.5",