diff --git a/src/tabpfn/architectures/tabpfn_v2_5.py b/src/tabpfn/architectures/tabpfn_v2_5.py
index a16a5f0aa..1d74d6cd6 100644
--- a/src/tabpfn/architectures/tabpfn_v2_5.py
+++ b/src/tabpfn/architectures/tabpfn_v2_5.py
@@ -568,8 +568,6 @@ def __init__(
             self.input_size // 4, self.input_size
         )
         self._do_encoder_nan_check = True
-        # TODO(Phil): This is here to not fail the memory computation. We should make
-        # this a proper API.
         self.emsize = config.emsize
 
     @property
diff --git a/src/tabpfn/architectures/tabpfn_v2_6.py b/src/tabpfn/architectures/tabpfn_v2_6.py
index 927d8e8b5..4b1e9c360 100644
--- a/src/tabpfn/architectures/tabpfn_v2_6.py
+++ b/src/tabpfn/architectures/tabpfn_v2_6.py
@@ -579,8 +579,6 @@ def __init__(
             self.input_size // 4, self.input_size
         )
         self._do_encoder_nan_check = True
-        # TODO(Phil): This is here to not fail the memory computation. We should make
-        # this a proper API.
         self.emsize = config.emsize
 
     @property
diff --git a/src/tabpfn/architectures/tabpfn_v3.py b/src/tabpfn/architectures/tabpfn_v3.py
index a3db94848..e36aff24e 100644
--- a/src/tabpfn/architectures/tabpfn_v3.py
+++ b/src/tabpfn/architectures/tabpfn_v3.py
@@ -269,8 +269,9 @@ def cache_size_mb(self) -> int:
     SDPBackend.FLASH_ATTENTION,
     SDPBackend.EFFICIENT_ATTENTION,
     SDPBackend.CUDNN_ATTENTION,
+    SDPBackend.MATH,  # fallback for older GPUs or unsupported configurations
 ]
-_SDPA_BACKENDS_CPU = [*_SDPA_BACKENDS, SDPBackend.MATH]
+_SDPA_BACKENDS_CPU = [*_SDPA_BACKENDS]
 
 
 # ---------------------------------------------------------------------------
diff --git a/src/tabpfn/model_loading.py b/src/tabpfn/model_loading.py
index ed325b075..4237b29f9 100644
--- a/src/tabpfn/model_loading.py
+++ b/src/tabpfn/model_loading.py
@@ -169,22 +169,22 @@ def get_regressor_v2_6(cls) -> ModelSource:  # noqa: D102
     @classmethod
     def get_classifier_v3(cls) -> ModelSource:  # noqa: D102
         filenames = [
-            "tabpfn-v3-classifier-20260506.ckpt",
+            "tabpfn-v3-classifier-v3_default.ckpt",
         ]
         return cls(
             repo_id="Prior-Labs/tabpfn_3",
-            default_filename="tabpfn-v3-classifier-20260506.ckpt",
+            default_filename="tabpfn-v3-classifier-v3_default.ckpt",
             filenames=filenames,
         )
 
     @classmethod
     def get_regressor_v3(cls) -> ModelSource:  # noqa: D102
         filenames = [
-            "tabpfn-v3-regressor-20260506.ckpt",
+            "tabpfn-v3-regressor-v3_default.ckpt",
         ]
         return cls(
             repo_id="Prior-Labs/tabpfn_3",
-            default_filename="tabpfn-v3-regressor-20260506.ckpt",
+            default_filename="tabpfn-v3-regressor-v3_default.ckpt",
             filenames=filenames,
         )
 
@@ -499,13 +499,14 @@ def _download_model(
         ModelVersion.V2_6: "tabpfn_2_6",
         ModelVersion.V3: "tabpfn_3",
     }
-    if version in _HF_REPOS:
-        try:
-            from tabpfn.browser_auth import ensure_license_accepted  # noqa: PLC0415
+    # Skip license check for now until tabpfn_3 is public
+    # if version in _HF_REPOS:
+    #     try:
+    #         from tabpfn.browser_auth import ensure_license_accepted
 
-            ensure_license_accepted(hf_repo_id=_HF_REPOS[version])
-        except Exception as e:  # noqa: BLE001
-            return [e]
+    #         ensure_license_accepted(hf_repo_id=_HF_REPOS[version])
+    #     except Exception as e:
+    #         return [e]
 
     try:
         model_source = _get_model_source(version, ModelType(which))
@@ -975,10 +976,10 @@ def load_model(
     )
 
     if "test_targets_MB" in inspect.signature(model.forward).parameters:
-        # The model computes the loss internally. Support for this was only added after
-        # v2.5, so we can safely assume that the inference config is stored in the
-        # checkpoint.
-        model.load_state_dict(full_state)
+        # The model computes the loss internally. Strip criterion keys that
+        # save_tabpfn_model may have written so load_state_dict doesn't reject them.
+        model_state = {k: v for k, v in full_state.items() if "criterion." not in k}
+        model.load_state_dict(model_state)
         model.eval()
         inference_config = InferenceConfig(
             **_rename_old_inference_config_keys(checkpoint["inference_config"])
@@ -1102,6 +1103,7 @@ def save_tabpfn_model(
             "state_dict": state_dict,
             "config": asdict(config),
             "architecture_name": architecture_name,
+            "inference_config": asdict(model.inference_config_),
         }
         if inference_config is not None:
             checkpoint["inference_config"] = asdict(inference_config)
@@ -1227,11 +1229,14 @@ def load_fitted_tabpfn_model(
 
 def _resolve_architecture_name(config: ArchitectureConfig) -> str:
     """Resolve the architecture name from the config."""
-    name = getattr(config, "name", "")
-    if "v3" in name:
+    from tabpfn.architectures.tabpfn_v2_5 import TabPFNV2p5Config  # noqa: PLC0415
+    from tabpfn.architectures.tabpfn_v2_6 import TabPFNV2p6Config  # noqa: PLC0415
+    from tabpfn.architectures.tabpfn_v3 import TabPFNV3Config  # noqa: PLC0415
+
+    if isinstance(config, TabPFNV3Config):
         return "tabpfn_v3"
-    if "2.6" in name:
+    if isinstance(config, TabPFNV2p6Config):
         return "tabpfn_v2_6"
-    if "2.5" in name:
+    if isinstance(config, TabPFNV2p5Config):
         return "tabpfn_v2_5"
     return "base"
diff --git a/src/tabpfn/settings.py b/src/tabpfn/settings.py
index 1efe46af5..4786e783f 100644
--- a/src/tabpfn/settings.py
+++ b/src/tabpfn/settings.py
@@ -35,7 +35,7 @@ class TabPFNSettings(BaseSettings):
         "If not set, uses platform-specific user cache directory.",
     )
     model_version: ModelVersion = Field(
-        default=ModelVersion.V2_6,
+        default=ModelVersion.V3,
         description="The version of the TabPFN model to use by default.",
     )
 
diff --git a/tests/conftest.py b/tests/conftest.py
index e802861dd..cc1308147 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -16,8 +16,6 @@
 import pytest
 import torch
 
-from tabpfn.model_loading import ModelSource, get_cache_dir
-
 
 def pytest_configure(config: pytest.Config) -> None:  # noqa: ARG001
     """Configure pytest with global settings."""
@@ -31,13 +29,3 @@ def set_global_seed() -> None:
     torch.manual_seed(seed)
     np.random.seed(seed)  # noqa: NPY002
     random.seed(seed)
-
-
-def is_v3_classifier_in_cache() -> bool:
-    cache_dir = get_cache_dir()
-    return (cache_dir / ModelSource.get_classifier_v3().default_filename).exists()
-
-
-def is_v3_regressor_in_cache() -> bool:
-    cache_dir = get_cache_dir()
-    return (cache_dir / ModelSource.get_regressor_v3().default_filename).exists()
diff --git a/tests/reference_predictions/darwin_arm64/classifier_iris_dataset.json b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices_v2.6.json
similarity index 100%
rename from tests/reference_predictions/darwin_arm64/classifier_iris_dataset.json
rename to tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices_v2.6.json
diff --git a/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices_v3.json b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices_v3.json
new file mode 100644
index 000000000..bdccf809b
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices_v3.json
@@ -0,0 +1,17 @@
+[
+  [
+    0.9999734163284302,
+    1.3569095244747587e-05,
+    1.2983076885575429e-05
+  ],
+  [
+    0.00036331909359432757,
+    0.9991976618766785,
+    0.0004389923997223377
+  ],
+  [
+    2.672282062121667e-05,
+    0.0005141739966347814,
+    0.9994590878486633
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices.json b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_v2.6.json
similarity index 100%
rename from tests/reference_predictions/darwin_arm64/classifier_iris_dataset_several_devices.json
rename to tests/reference_predictions/darwin_arm64/classifier_iris_dataset_v2.6.json
diff --git a/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_v3.json b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_v3.json
new file mode 100644
index 000000000..bdccf809b
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_iris_dataset_v3.json
@@ -0,0 +1,17 @@
+[
+  [
+    0.9999734163284302,
+    1.3569095244747587e-05,
+    1.2983076885575429e-05
+  ],
+  [
+    0.00036331909359432757,
+    0.9991976618766785,
+    0.0004389923997223377
+  ],
+  [
+    2.672282062121667e-05,
+    0.0005141739966347814,
+    0.9994590878486633
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v2.6.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v2.6.json
new file mode 100644
index 000000000..0eecbfbe6
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v2.6.json
@@ -0,0 +1,14 @@
+[
+  [
+    0.745369017124176,
+    0.2546309530735016
+  ],
+  [
+    0.42508718371391296,
+    0.5749127864837646
+  ],
+  [
+    0.6312955617904663,
+    0.3687044084072113
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v3.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v3.json
new file mode 100644
index 000000000..220fa9c68
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_3_estimators_v3.json
@@ -0,0 +1,14 @@
+[
+  [
+    0.6735507845878601,
+    0.3264492154121399
+  ],
+  [
+    0.44950684905052185,
+    0.5504931807518005
+  ],
+  [
+    0.6495934128761292,
+    0.35040655732154846
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_5_estimators.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_5_estimators.json
deleted file mode 100644
index e6ac22e13..000000000
--- a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_5_estimators.json
+++ /dev/null
@@ -1,14 +0,0 @@
-[
-  [
-    0.6925381422042847,
-    0.3074618875980377
-  ],
-  [
-    0.4254949390888214,
-    0.574505090713501
-  ],
-  [
-    0.6501684188842773,
-    0.34983158111572266
-  ]
-]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input_v2.6.json
similarity index 100%
rename from tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input.json
rename to tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input_v2.6.json
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input_v3.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input_v3.json
new file mode 100644
index 000000000..bd4ff24a3
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_differentiable_input_v3.json
@@ -0,0 +1,14 @@
+[
+  [
+    0.7227111458778381,
+    0.27728888392448425
+  ],
+  [
+    0.47503453493118286,
+    0.5249655246734619
+  ],
+  [
+    0.4784924387931824,
+    0.5215075612068176
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_v3.json b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_v3.json
new file mode 100644
index 000000000..ea0801da0
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/classifier_tiny_dataset_v3.json
@@ -0,0 +1,14 @@
+[
+  [
+    0.678566575050354,
+    0.3214334547519684
+  ],
+  [
+    0.46058961749076843,
+    0.5394103527069092
+  ],
+  [
+    0.6697312593460083,
+    0.3302687108516693
+  ]
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices.json b/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices_v2.6.json
similarity index 100%
rename from tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices.json
rename to tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices_v2.6.json
diff --git a/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices_v3.json b/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices_v3.json
new file mode 100644
index 000000000..4a8c68fbf
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_several_devices_v3.json
@@ -0,0 +1,5 @@
+[
+  5.081058502197266,
+  4.23805046081543,
+  4.6899261474609375
+]
\ No newline at end of file
diff --git a/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_v3.json b/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_v3.json
new file mode 100644
index 000000000..4a8c68fbf
--- /dev/null
+++ b/tests/reference_predictions/darwin_arm64/regressor_tiny_dataset_v3.json
@@ -0,0 +1,5 @@
+[
+  5.081058502197266,
+  4.23805046081543,
+  4.6899261474609375
+]
\ No newline at end of file
diff --git a/tests/test_classifier_interface.py b/tests/test_classifier_interface.py
index da5e66c9c..93d293f7c 100644
--- a/tests/test_classifier_interface.py
+++ b/tests/test_classifier_interface.py
@@ -36,7 +36,6 @@
 from tabpfn.preprocessing import PreprocessorConfig
 from tabpfn.utils import infer_devices
 
-from .conftest import is_v3_classifier_in_cache
 from .utils import (
     get_pytest_devices,
     is_cpu_float16_supported,
@@ -64,7 +63,11 @@ def X_y() -> tuple[np.ndarray, np.ndarray]:
     )
 
 
-model_sources = [ModelSource.get_classifier_v2(), ModelSource.get_classifier_v2_5()]
+model_sources = [
+    ModelSource.get_classifier_v2(),
+    ModelSource.get_classifier_v2_5(),
+    ModelSource.get_classifier_v3(),
+]
 fit_modes = ["low_memory", "fit_preprocessors"]
 
 
@@ -525,8 +528,9 @@ def test_balance_probabilities_alters_proba_output() -> None:
     )
 
 
-# Only v2 and 2.5 support the KV cache at the moment.
-@pytest.mark.parametrize("model_version", [ModelVersion.V2, ModelVersion.V2_5])
+@pytest.mark.parametrize(
+    "model_version", [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V3]
+)
 # Disable MPS as it doesn't support float64.
 @pytest.mark.parametrize("device", [d for d in get_pytest_devices() if d != "mps"])
 def test__fit_preprocessors_and_with_cache_produce_equal_results(
@@ -564,8 +568,6 @@ def test__fit_preprocessors_and_with_cache_produce_equal_results(
 def test__fit_preprocessors_and_low_memory_produce_equal_results(
     X_y: tuple[np.ndarray, np.ndarray], model_version: ModelVersion, device: str
 ) -> None:
-    if model_version == ModelVersion.V3 and not is_v3_classifier_in_cache():
-        pytest.skip("V3 classifier model not in cache; skipping V3-specific test.")
     kwargs = {
         "version": model_version,
         "n_estimators": 2,
@@ -596,9 +598,6 @@ def test__fit_preprocessors_and_low_memory_produce_equal_results(
 def test__fit_and_predict__on_demo_dataset__accuracy_reasonable(
     model_version: ModelVersion,
 ) -> None:
-    if model_version == ModelVersion.V3 and not is_v3_classifier_in_cache():
-        pytest.skip("V3 classifier model not in cache.")
-
     X, y = sklearn.datasets.load_iris(return_X_y=True)
     model = TabPFNClassifier.create_default_for_version(
         version=model_version, random_state=0
diff --git a/tests/test_consistency.py b/tests/test_consistency.py
index 3fdcf3682..03e4f3700 100644
--- a/tests/test_consistency.py
+++ b/tests/test_consistency.py
@@ -117,7 +117,12 @@ class _ConsistencyCase:
                 **DEFAULT_CONFIG,
             ),
         )
-        for version in [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V2_6]
+        for version in [
+            ModelVersion.V2,
+            ModelVersion.V2_5,
+            ModelVersion.V2_6,
+            ModelVersion.V3,
+        ]
     },
     **{
         f"regressor_tiny_dataset_{version.value}": _ConsistencyCase(
@@ -128,21 +133,38 @@ class _ConsistencyCase:
                 **DEFAULT_CONFIG,
             ),
         )
-        for version in [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V2_6]
+        for version in [
+            ModelVersion.V2,
+            ModelVersion.V2_5,
+            ModelVersion.V2_6,
+            ModelVersion.V3,
+        ]
     },
-    "classifier_tiny_dataset_differentiable_input": _ConsistencyCase(
+    "classifier_tiny_dataset_differentiable_input_v2.6": _ConsistencyCase(
         data=lambda: _to_tensors(_get_tiny_classification_data()),
         model=lambda: TabPFNClassifier.create_default_for_version(
             version=ModelVersion.V2_6, **DEFAULT_CONFIG, differentiable_input=True
         ),
     ),
-    "classifier_iris_dataset": _ConsistencyCase(
+    "classifier_tiny_dataset_differentiable_input_v3": _ConsistencyCase(
+        data=lambda: _to_tensors(_get_tiny_classification_data()),
+        model=lambda: TabPFNClassifier.create_default_for_version(
+            version=ModelVersion.V3, **DEFAULT_CONFIG, differentiable_input=True
+        ),
+    ),
+    "classifier_iris_dataset_v2.6": _ConsistencyCase(
         data=_get_iris_multiclass_data,
         model=lambda: TabPFNClassifier.create_default_for_version(
             version=ModelVersion.V2_6, **DEFAULT_CONFIG
         ),
     ),
-    "regressor_tiny_dataset_several_devices": _ConsistencyCase(
+    "classifier_iris_dataset_v3": _ConsistencyCase(
+        data=_get_iris_multiclass_data,
+        model=lambda: TabPFNClassifier.create_default_for_version(
+            version=ModelVersion.V3, **DEFAULT_CONFIG
+        ),
+    ),
+    "regressor_tiny_dataset_several_devices_v2.6": _ConsistencyCase(
         data=_get_tiny_regression_data,
         model=lambda: _add_extra_devices(
             TabPFNRegressor.create_default_for_version(
@@ -150,7 +172,15 @@ class _ConsistencyCase:
             )
         ),
     ),
-    "classifier_iris_dataset_several_devices": _ConsistencyCase(
+    "regressor_tiny_dataset_several_devices_v3": _ConsistencyCase(
+        data=_get_tiny_regression_data,
+        model=lambda: _add_extra_devices(
+            TabPFNRegressor.create_default_for_version(
+                version=ModelVersion.V3, **DEFAULT_CONFIG
+            )
+        ),
+    ),
+    "classifier_iris_dataset_several_devices_v2.6": _ConsistencyCase(
         data=_get_iris_multiclass_data,
         model=lambda: _add_extra_devices(
             TabPFNClassifier.create_default_for_version(
@@ -158,10 +188,24 @@ class _ConsistencyCase:
             )
         ),
     ),
-    "classifier_tiny_dataset_5_estimators": _ConsistencyCase(
+    "classifier_iris_dataset_several_devices_v3": _ConsistencyCase(
+        data=_get_iris_multiclass_data,
+        model=lambda: _add_extra_devices(
+            TabPFNClassifier.create_default_for_version(
+                version=ModelVersion.V3, **DEFAULT_CONFIG
+            )
+        ),
+    ),
+    "classifier_tiny_dataset_3_estimators_v2.6": _ConsistencyCase(
+        data=_get_tiny_classification_data,
+        model=lambda: TabPFNClassifier.create_default_for_version(
+            version=ModelVersion.V2_6, **DEFAULT_CONFIG | {"n_estimators": 3}
+        ),
+    ),
+    "classifier_tiny_dataset_3_estimators_v3": _ConsistencyCase(
         data=_get_tiny_classification_data,
         model=lambda: TabPFNClassifier.create_default_for_version(
-            version=ModelVersion.V2_6, **DEFAULT_CONFIG | {"n_estimators": 5}
+            version=ModelVersion.V3, **DEFAULT_CONFIG | {"n_estimators": 3}
         ),
     ),
 }
diff --git a/tests/test_estimators.py b/tests/test_estimators.py
index de9b225c6..192bf48b9 100644
--- a/tests/test_estimators.py
+++ b/tests/test_estimators.py
@@ -127,7 +127,8 @@ def test__to__after_fit__no_tensors_left_on_old_device(
 @pytest.mark.parametrize("estimator_class", [TabPFNRegressor, TabPFNClassifier])
 @pytest.mark.parametrize("fit_mode", ["fit_preprocessors", "low_memory"])
 @pytest.mark.parametrize(
-    "model_version", [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V2_6]
+    "model_version",
+    [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V2_6, ModelVersion.V3],
 )
 def test__to__after_fit_and_predict__no_tensors_left_on_old_device(
     estimator_class: type[TabPFNClassifier] | type[TabPFNRegressor],
diff --git a/tests/test_inference.py b/tests/test_inference.py
index 89b7a36fc..b652fd952 100644
--- a/tests/test_inference.py
+++ b/tests/test_inference.py
@@ -76,6 +76,7 @@ def forward(
         return x.sum(-2, keepdim=True).sum(-1, keepdim=True).reshape(-1, test_rows)
 
     @property
+    @override
     def embedding_dim(self) -> int:
         return 2
 
@@ -118,6 +119,7 @@ def forward(
         return x.sum(-2, keepdim=True).sum(-1, keepdim=True).reshape(-1, test_rows)
 
     @property
+    @override
     def embedding_dim(self) -> int:
         return 2
 
diff --git a/tests/test_regressor_interface.py b/tests/test_regressor_interface.py
index 0e3c046a0..1f6b912f2 100644
--- a/tests/test_regressor_interface.py
+++ b/tests/test_regressor_interface.py
@@ -30,7 +30,6 @@
 from tabpfn.settings import settings
 from tabpfn.utils import infer_devices
 
-from .conftest import is_v3_regressor_in_cache
 from .utils import (
     get_pytest_devices,
     is_cpu_float16_supported,
@@ -41,7 +40,11 @@
 devices = get_pytest_devices()
 
 
-model_sources = [ModelSource.get_regressor_v2(), ModelSource.get_regressor_v2_5()]
+model_sources = [
+    ModelSource.get_regressor_v2(),
+    ModelSource.get_regressor_v2_5(),
+    ModelSource.get_regressor_v3(),
+]
 fit_modes = ["low_memory", "fit_preprocessors"]
 
 
@@ -280,8 +283,9 @@ def test__fit_predict__specify_inference_config__outputs_correct_shape(
     assert model.predict(X).shape == (X.shape[0],)
 
 
-# Only v2 and 2.5 support the KV cache at the moment.
-@pytest.mark.parametrize("model_version", [ModelVersion.V2, ModelVersion.V2_5])
+@pytest.mark.parametrize(
+    "model_version", [ModelVersion.V2, ModelVersion.V2_5, ModelVersion.V3]
+)
 # Disable MPS as it doesn't support float64.
 @pytest.mark.parametrize("device", [d for d in get_pytest_devices() if d != "mps"])
 def test__fit_preprocessors_and_with_cache_produce_equal_results(
@@ -317,8 +321,6 @@ def test__fit_preprocessors_and_with_cache_produce_equal_results(
 def test__fit_preprocessors_and_low_memory_produce_equal_results(
     X_y: tuple[np.ndarray, np.ndarray], model_version: ModelVersion, device: str
 ) -> None:
-    if model_version == ModelVersion.V3 and not is_v3_regressor_in_cache():
-        pytest.skip("V3 regressor model not in cache; skipping V3-specific test.")
     kwargs = {
         "version": model_version,
         "n_estimators": 2,
@@ -345,9 +347,6 @@ def test__fit_preprocessors_and_low_memory_produce_equal_results(
 def test__fit_and_predict__on_demo_dataset__r2_reasonable(
     model_version: ModelVersion,
 ) -> None:
-    if model_version == ModelVersion.V3 and not is_v3_regressor_in_cache():
-        pytest.skip("V3 regressor model not in cache.")
-
     X, y = sklearn.datasets.make_friedman1(n_samples=200, noise=0.1, random_state=0)
     model = TabPFNRegressor.create_default_for_version(
         version=model_version, random_state=0