diff --git a/routstr/core/main.py b/routstr/core/main.py
index 4defd7cb..54004699 100644
--- a/routstr/core/main.py
+++ b/routstr/core/main.py
@@ -148,12 +148,13 @@ async def lifespan(_: FastAPI) -> AsyncGenerator[None, None]:
 
         # await ensure_models_bootstrapped()
 
-        from ..proxy import get_upstreams
+        from ..proxy import get_upstreams, mark_warming_up
         from ..upstream.helpers import refresh_upstreams_models_periodically
 
         # Provider discovery hits every upstream's /models (plus the OpenRouter
         # catalog for unpriced models), so keep it off the readiness path: the
         # app serves as soon as the DB is up and fills its model maps after.
+        mark_warming_up()
         bootstrap_task = asyncio.create_task(_bootstrap_providers_and_pricing())
 
         btc_price_task = asyncio.create_task(update_prices_periodically())
diff --git a/routstr/proxy.py b/routstr/proxy.py
index a2cf9492..7c0cf908 100644
--- a/routstr/proxy.py
+++ b/routstr/proxy.py
@@ -25,6 +25,7 @@ from .core.db import (
 )
 from .core.error_scope import (
     ERROR_SCOPE_HEADER,
+    ERROR_SCOPE_NODE,
     ERROR_SCOPE_UPSTREAM,
     UPSTREAM_ERROR_STATUS,
     UPSTREAM_UNAVAILABLE,
@@ -62,8 +63,14 @@ from .upstream.request_correction import correct_request, extract_error_message
 logger = get_logger(__name__)
 
 MODEL_PATH_HEADER = "x-routstr-model-path"
+MODELS_WARMING_UP = "MODELS_WARMING_UP"
+WARMING_UP_RETRY_AFTER_SECONDS = 2
 proxy_router = APIRouter()
 
+# Startup loads providers in the background, so until that first load finishes
+# an unknown model may just not be loaded yet.
+_warming_up = False
+
 _upstreams: list[BaseUpstreamProvider] = []
 _provider_map: dict[
     str, list[tuple[Model, BaseUpstreamProvider]]
@@ -80,10 +87,23 @@ async def _finish_read_transaction(session: AsyncSession) -> None:
 
 async def initialize_upstreams() -> None:
     """Initialize upstream providers from database during application startup."""
-    global _upstreams
-    _upstreams = await init_upstreams()
-    logger.info(f"Initialized {len(_upstreams)} upstream providers")
-    await refresh_model_maps()
+    global _upstreams, _warming_up
+    try:
+        _upstreams = await init_upstreams()
+        logger.info(f"Initialized {len(_upstreams)} upstream providers")
+        await refresh_model_maps()
+    finally:
+        _warming_up = False
+
+
+def mark_warming_up() -> None:
+    """Answer unknown models with a retryable 503 until initialize_upstreams() ends."""
+    global _warming_up
+    _warming_up = True
+
+
+def is_warming_up() -> bool:
+    return _warming_up
 
 
 async def reinitialize_upstreams() -> None:
@@ -650,6 +670,17 @@ async def _proxy(
     candidates = get_candidates(model_id)
 
     if not candidates:
+        if is_warming_up():
+            response = create_error_response(
+                "service_unavailable",
+                "Models are still loading, retry shortly",
+                503,
+                request=request,
+                code=MODELS_WARMING_UP,
+                error_scope=ERROR_SCOPE_NODE,
+            )
+            response.headers["Retry-After"] = str(WARMING_UP_RETRY_AFTER_SECONDS)
+            return response
         return create_error_response(
             "invalid_model", f"Model '{model_id}' not found", 400, request=request
         )
diff --git a/tests/unit/test_models_warming_up.py b/tests/unit/test_models_warming_up.py
new file mode 100644
index 00000000..99f286e7
--- /dev/null
+++ b/tests/unit/test_models_warming_up.py
@@ -0,0 +1,58 @@
+"""Unknown models get a retryable 503 while startup is still loading providers."""
+
+import json
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from routstr import proxy as proxy_module
+from routstr.core.error_scope import ERROR_SCOPE_HEADER, ERROR_SCOPE_NODE
+
+from .test_model_path_routing import _make_request, _run_proxy
+
+
+@pytest.fixture(autouse=True)
+def _reset_warming_up(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setattr(proxy_module, "_warming_up", False)
+
+
+def _chat_request() -> MagicMock:
+    return _make_request(
+        {"authorization": "Bearer sk-mpkey"},
+        json.dumps({"model": "not-loaded-yet"}).encode(),
+    )
+
+
+@pytest.mark.asyncio
+async def test_unknown_model_while_warming_up_is_retryable_503() -> None:
+    proxy_module.mark_warming_up()
+
+    response = await _run_proxy(_chat_request(), [])
+
+    assert response.status_code == 503
+    assert response.headers["Retry-After"] == "2"
+    assert response.headers[ERROR_SCOPE_HEADER] == ERROR_SCOPE_NODE
+    assert json.loads(bytes(response.body))["error"]["code"] == "MODELS_WARMING_UP"
+
+
+@pytest.mark.asyncio
+async def test_unknown_model_after_warm_up_is_invalid_model() -> None:
+    response = await _run_proxy(_chat_request(), [])
+
+    assert response.status_code == 400
+    assert json.loads(bytes(response.body))["error"]["type"] == "invalid_model"
+
+
+@pytest.mark.asyncio
+async def test_failed_initialization_still_ends_warm_up() -> None:
+    proxy_module.mark_warming_up()
+
+    with (
+        patch.object(
+            proxy_module, "init_upstreams", AsyncMock(side_effect=RuntimeError)
+        ),
+        pytest.raises(RuntimeError),
+    ):
+        await proxy_module.initialize_upstreams()
+
+    assert not proxy_module.is_warming_up()