diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 490ba27e..34e81487 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -6,36 +6,64 @@ on:
     paths:
       - "**.py"
       - "pyproject.toml"
+      - "uv.lock"
       - ".github/workflows/test.yml"
   pull_request:
     branches: [main]
     paths:
       - "**.py"
       - "pyproject.toml"
+      - "uv.lock"
       - ".github/workflows/test.yml"
 
+permissions:
+  contents: read
+
 jobs:
-  test:
-    name: Renderers
+  unit:
+    name: Offline (Python ${{ matrix.python-version }})
     runs-on: ubuntu-latest
+    env:
+      HF_HUB_OFFLINE: "1"
+      TRANSFORMERS_OFFLINE: "1"
     strategy:
       fail-fast: false
       matrix:
         python-version: ["3.11", "3.12", "3.13", "3.14"]
-
     steps:
       - uses: actions/checkout@v4
-
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v6
+      - uses: actions/setup-python@v6
         with:
           python-version: ${{ matrix.python-version }}
+      - uses: astral-sh/setup-uv@v7
+      - run: uv sync --locked
+      - run: uv run --no-sync pytest tests/ -m "not model" -v
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v7
-
-      - name: Install dependencies
-        run: uv sync
-
-      - name: Run tests
-        run: uv run pytest tests/ -v
+  model:
+    name: Pinned ${{ matrix.suite }} parity
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - suite: text
+            selection: "model and not multimodal"
+          - suite: multimodal
+            selection: "model and multimodal"
+    env:
+      HF_HUB_DISABLE_PROGRESS_BARS: "1"
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.13"
+      - uses: astral-sh/setup-uv@v7
+      - uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cache/huggingface/hub
+            ~/.cache/huggingface/modules
+          key: hf-renderers-${{ runner.os }}-${{ matrix.suite }}-${{ hashFiles('tests/model_assets.py') }}
+      - run: uv sync --locked
+      - run: uv run --no-sync pytest tests/ -m "${{ matrix.selection }}" -v
diff --git a/pyproject.toml b/pyproject.toml
index 09daae21..3fac356b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -139,3 +139,9 @@ invalid-key = "warn"
 unsupported-operator = "warn"
 invalid-return-type = "warn"
 invalid-assignment = "warn"
+
+[tool.pytest.ini_options]
+markers = [
+    "model: requires real tokenizer or processor assets at pinned revisions",
+    "multimodal: exercises real image/audio/video processors or model sidecars",
+]
diff --git a/renderers/base.py b/renderers/base.py
index f3418841..d7ce49ad 100644
--- a/renderers/base.py
+++ b/renderers/base.py
@@ -1313,13 +1313,14 @@ def _load_tokenizer_via_auto(model_name_or_path: str, **kwargs) -> Any:
         return tok
 
 
-def load_tokenizer(model_name_or_path: str):
+def load_tokenizer(model_name_or_path: str, *, revision: str | None = None):
     """Load a tokenizer with the renderers-package security policy.
 
     Default ``trust_remote_code=False``. Models listed in
     ``TRUSTED_REVISIONS`` (Moonshot Kimi-K2 family) load with
     ``trust_remote_code=True`` AND a pinned ``revision=<sha>`` so
     transformers only executes the reviewed commit's tokenizer Python.
+    ``revision`` pins ordinary assets without overriding reviewed remote code.
 
     ``AutoTokenizer.from_pretrained`` eagerly builds the model config to
     resolve the tokenizer class. If that construction raises on a
@@ -1338,6 +1339,11 @@ def load_tokenizer(model_name_or_path: str):
         _require_transformers("Loading a tokenizer")
         load_name_or_path = _tokenizer_source_for(model_name_or_path)
         kwargs = _tokenizer_load_kwargs(load_name_or_path)
+        if revision is not None:
+            trusted_revision = kwargs.get("revision")
+            if trusted_revision is not None and revision != trusted_revision:
+                raise ValueError("Cannot override the reviewed remote-code revision")
+            kwargs["revision"] = revision
         tok = _load_tokenizer_via_auto(load_name_or_path, **kwargs)
         return _preserve_requested_tokenizer_name(
             tok,
diff --git a/tests/conftest.py b/tests/conftest.py
index 7b9f9732..cce8b596 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -12,9 +12,10 @@
 
 import pytest
 from parity import models_for
+
 from renderers import create_renderer
-from renderers.base import load_tokenizer
 from renderers.configs import config_from_name
+from tests.model_assets import load_test_tokenizer
 
 # Backwards-compatible view used by documentation and a few invariant tests.
 RENDERER_MODELS = [(case.model, case.renderer) for case in models_for("shared")]
@@ -33,7 +34,7 @@ def release_shared_models():
 def _load(model_name: str, renderer_name: str):
     key = f"{model_name}:{renderer_name}"
     if key not in _cache:
-        tokenizer = load_tokenizer(model_name)
+        tokenizer = load_test_tokenizer(model_name)
         renderer = create_renderer(tokenizer, config_from_name(renderer_name))
         _cache[key] = (tokenizer, renderer)
     return _cache[key]
@@ -54,7 +55,10 @@ def pytest_generate_tests(metafunc):
         cases = models_for(suite)
         metafunc.parametrize(
             "model_name,renderer_name",
-            [(case.model, case.renderer) for case in cases],
+            [
+                pytest.param(case.model, case.renderer, marks=pytest.mark.model)
+                for case in cases
+            ],
             ids=[case.model for case in cases],
         )
 
diff --git a/tests/model_assets.py b/tests/model_assets.py
new file mode 100644
index 00000000..bdac1f6c
--- /dev/null
+++ b/tests/model_assets.py
@@ -0,0 +1,94 @@
+"""Immutable test assets; Llama revisions refer to the audited tokenizer mirrors."""
+
+from pathlib import Path
+
+from renderers.base import load_tokenizer
+
+MODEL_REVISIONS = {
+    "MiniMaxAI/MiniMax-M2.5": "2515a3ea0afec1dd22d0c7794b3a91557804d195",
+    "PrimeIntellect/Qwen3-0.6B": "15aba67c8ec68aeac96f57ba7e31d373564ee03c",
+    "PrimeIntellect/Qwen3-1.7B": "61a79cd101170642a653d1d44d18f69f88f54e08",
+    "Qwen/Qwen2.5-0.5B-Instruct": "7ae557604adf67be50417f59c2c2f167def9a775",
+    "Qwen/Qwen3-0.6B": "c1899de289a04d12100db370d81485cdf75e47ca",
+    "Qwen/Qwen3-4B": "1cfa9a7208912126459214e8b04321603b3df60c",
+    "Qwen/Qwen3-8B": "b968826d9c46dd6066d109eabc6255188de91218",
+    "Qwen/Qwen3-VL-30B-A3B-Instruct": "9c4b90e1e4ba969fd3b5378b57d966d725f1b86c",
+    "Qwen/Qwen3-VL-4B-Instruct": "ebb281ec70b05090aa6165b016eac8ec08e71b17",
+    "Qwen/Qwen3-VL-8B-Instruct": "0c351dd01ed87e9c1b53cbc748cba10e6187ff3b",
+    "Qwen/Qwen3.5-0.8B": "2fc06364715b967f1860aea9cf38778875588b17",
+    "Qwen/Qwen3.5-122B-A10B": "dc4d348443bc740c68e2d77492492c11606384d5",
+    "Qwen/Qwen3.5-2B": "15852e8c16360a2fea060d615a32b45270f8a8fc",
+    "Qwen/Qwen3.5-35B-A3B": "59d61f3ce65a6d9863b86d2e96597125219dc754",
+    "Qwen/Qwen3.5-397B-A17B": "8472618112abcbd45acbcdc58436aff4233c23f7",
+    "Qwen/Qwen3.5-4B": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "Qwen/Qwen3.5-9B": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
+    "Qwen/Qwen3.6-35B-A3B": "995ad96eacd98c81ed38be0c5b274b04031597b0",
+    "Qwen/Qwen3.8-27B": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
+    "Qwen/Qwen3.8-Flash-Next": "de4b8e4d43b917e7706784d8bb445c9af86a3540",
+    "THUDM/GLM-4.5-Air": "a24ceef6ce4f3536971efe9b778bdaa1bab18daa",
+    "deepseek-ai/DeepSeek-R1": "56d4cbbb4d29f4355bab4b9a39ccb717a14ad5ad",
+    "deepseek-ai/DeepSeek-V3": "e815299b0bcbac849fa540c768ef21845365c9eb",
+    "deepseek-ai/DeepSeek-V4-Flash-0731": "7872f01b1d1fe23eabc4c98b48bffcef5a386062",
+    "google/gemma-4-26B-A4B-it": "4d7ae4984b7db7de8f8457170b3f1a419ee76d52",
+    "google/gemma-4-31B-it": "842da3794eaa0b77d5f08bae87a17459d91ff475",
+    "google/gemma-4-E2B-it": "3e22461f65e89153144f8adb70e3b8c2cc9845a7",
+    "google/gemma-4-E4B-it": "ee0ef6023621cff504d758262d4e04895a5af4a2",
+    "meta-llama/Llama-3.2-1B-Instruct": "5a8abab4a5d6f164389b1079fb721cfab8d7126c",
+    "meta-llama/Llama-3.2-3B-Instruct": "006f5dcd1393c3add266de40994ba96225e9689d",
+    "meta-models/Muse-Glimmer-30B": "cec94fcfd6512f2cfe708138c32533b346ea4f69",
+    "moonshotai/Kimi-K2-Instruct": "fd1984e2b7a3350dbf7305fe73a4ede25c14de50",
+    "moonshotai/Kimi-K2.5": "4d01dfe0332d63057c186e0b262165819efb6611",
+    "moonshotai/Kimi-K2.6": "2755962d07cb42aa2d988a35bcb65cd4a9c2de82",
+    "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": "bf77c3174f68ad409e1c2aa60daeb46e32d1c606",
+    "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16": "2dc98e2afe4face0e4ce40972a915c45368bd34a",
+    "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16": "77df655d5e9f8362164ed14dd8b48f8bce657498",
+    "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16": "a9904d24bcc1d289a1950fa9d2b978c47cf903b9",
+    "openai/gpt-oss-120b": "b5c939de8f754692c1647ca79fbf85e8c1e70f8a",
+    "openai/gpt-oss-20b": "6cee5e81ee83917806bbde320786a8fb61efebee",
+    "poolside/Laguna-M.1": "2bf8a4ab6d2e8e1e0d92a41318532a0fee1640d6",
+    "poolside/Laguna-S-2.1": "0f573140834b11cfac0c2af97a101a7a69a13e22",
+    "poolside/Laguna-XS-2.1": "c5f36269bbdbd3f27fddc9a9f9dbae0cf2cf57db",
+    "poolside/Laguna-XS.2": "a397bde04501591d0ac3d74a5d9d0a4ec298ffb8",
+    "tencent/Hy3": "a960ebc3da325ba167f069f76c41eb62c9280d22",
+    "thinkingmachines/Inkling": "828496eeae4c243ff1a22f7f28ff83694f2f7bc9",
+    "thinkingmachines/Inkling-Small": "8cc5877b44d343f88b92086aa1fb72897950f06a",
+    "zai-org/GLM-4.7-Flash": "7dd20894a642a0aa287e9827cb1a1f7f91386b67",
+    "zai-org/GLM-5": "c183ef8c61faee82855eca1ed9bb3a9a7ce3b0b2",
+    "zai-org/GLM-5.1": "26e1bd6e011feb778d25ae34b09b07074139d92d",
+    "zai-org/GLM-5.3": "ba32f8c19c8738e2c57c71798c601d477a7bf844",
+    "zai-org/GLM-5.3-BF16": "abf42e6c0e53bc29a875511301e7b7fcd220315b",
+    "zai-org/GLM-5.3-Flash-BF16": "f9fed3bbc82291073373d67452e9f25317ef7577",
+}
+
+
+def model_revision(model: str) -> str:
+    # The optional private VLM fixture can only resolve an existing local snapshot.
+    if model == "nvidia/NVIDIA-Nemotron-3.5-Super-EA-09112026":
+        from huggingface_hub import try_to_load_from_cache
+
+        path = try_to_load_from_cache(model, "tokenizer_config.json")
+        if isinstance(path, str):
+            return Path(path).parent.name
+    try:
+        return MODEL_REVISIONS[model]
+    except KeyError as exc:
+        raise KeyError(f"No immutable test revision registered for {model!r}") from exc
+
+
+def load_test_tokenizer(model: str):
+    return load_tokenizer(model, revision=model_revision(model))
+
+
+def load_test_processor(model: str):
+    from transformers import AutoProcessor
+
+    return AutoProcessor.from_pretrained(
+        model,
+        revision=model_revision(model),
+        trust_remote_code=model
+        in {
+            "moonshotai/Kimi-K2.5",
+            "moonshotai/Kimi-K2.6",
+            "nvidia/NVIDIA-Nemotron-3.5-Super-EA-09112026",
+        },
+    )
diff --git a/tests/test_bridge.py b/tests/test_bridge.py
index af5eaf99..a49374e1 100644
--- a/tests/test_bridge.py
+++ b/tests/test_bridge.py
@@ -27,8 +27,8 @@
 @lru_cache(maxsize=None)
 def _load(model_name: str, renderer_name: str):
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
     from renderers.configs import config_from_name
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer(model_name)
     return tok, create_renderer(tok, config_from_name(renderer_name))
@@ -38,7 +38,10 @@ def pytest_generate_tests(metafunc):
     if "br_model" in metafunc.fixturenames:
         metafunc.parametrize(
             "br_model,br_renderer_name",
-            _BRIDGE_MODELS,
+            [
+                pytest.param(model, renderer, marks=pytest.mark.model)
+                for model, renderer in _BRIDGE_MODELS
+            ],
             ids=[m for m, _ in _BRIDGE_MODELS],
         )
 
@@ -223,6 +226,7 @@ def test_qwen3_renders_openai_text_content_parts(br_renderer, br_tokenizer, br_m
     assert text in br_tokenizer.decode(list_ids, skip_special_tokens=False)
 
 
+@pytest.mark.model
 def test_bridge_declines_across_user_query_when_template_drops_thinking():
     """Qwen3's template drops a past block's thinking once a new user turn
     arrives. The resolved ``tool_cycle`` bridge policy therefore treats a
@@ -237,8 +241,8 @@ def test_bridge_declines_across_user_query_when_template_drops_thinking():
       - no prior thinking + new user query -> decline; no marker lookback.
     """
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
     from renderers.configs import Qwen3RendererConfig
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer("Qwen/Qwen3-8B")
     im_end = tok.convert_tokens_to_ids("<|im_end|>")
@@ -326,14 +330,15 @@ def test_bridge_declines_across_user_turn_when_thinking_present(br_renderer, br_
     assert extended is not None, f"{br_model}: should still bridge within a tool cycle"
 
 
+@pytest.mark.model
 def test_bridge_keeps_thinking_when_history_kwarg_disables_truncation():
     """GLM ``clear_thinking=False`` / Nemotron ``truncate_history_thinking=
     False`` keep all past thinking (the template doesn't strip it), so the
     bridge must NOT decline across a user turn — declining would re-render and
     re-tokenize model-sampled thinking bytes."""
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
     from renderers.configs import GLM5RendererConfig, Nemotron3RendererConfig
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     asst = [
         {"role": "assistant", "reasoning_content": "Let me think.", "content": "Hi"}
diff --git a/tests/test_client.py b/tests/test_client.py
index 3b53fe25..b85a6962 100644
--- a/tests/test_client.py
+++ b/tests/test_client.py
@@ -738,6 +738,8 @@ def test_generate_threads_prompt_attribution_through_prebuilt_prompt_path():
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.model
+@pytest.mark.multimodal
 @pytest.mark.parametrize(
     "model_id,renderer_class_path",
     [
@@ -762,11 +764,8 @@ def test_generate_serializes_multimodal_features_for_qwen_vl_family(
 
     import torch as _torch
 
-    from renderers.base import (
-        MultiModalData,
-        PlaceholderRange,
-        load_tokenizer,
-    )
+    from renderers.base import MultiModalData, PlaceholderRange
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     mod_name, cls_name = renderer_class_path.split(":")
     renderer_cls = getattr(importlib.import_module(mod_name), cls_name)
@@ -832,6 +831,8 @@ def test_generate_serializes_multimodal_features_for_qwen_vl_family(
         assert isinstance(item, str) and len(item) > 0
 
 
+@pytest.mark.model
+@pytest.mark.multimodal
 def test_generate_serializes_multimodal_features_for_gemma4():
     """Gemma 4's HF image positions are translated to vLLM's field name."""
     pytest.importorskip("torch")
@@ -839,8 +840,9 @@ def test_generate_serializes_multimodal_features_for_gemma4():
 
     import torch as _torch
 
-    from renderers.base import MultiModalData, PlaceholderRange, load_tokenizer
+    from renderers.base import MultiModalData, PlaceholderRange
     from renderers.gemma4 import Gemma4Renderer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     renderer = Gemma4Renderer(load_tokenizer("google/gemma-4-31B-it"))
     mm_data = MultiModalData(
diff --git a/tests/test_custom.py b/tests/test_custom.py
index c07da1b9..bc755207 100644
--- a/tests/test_custom.py
+++ b/tests/test_custom.py
@@ -169,8 +169,9 @@ def _emit_generation_prompt(self, emit_special) -> None:
 )
 
 
+@pytest.mark.model
 def test_readme_example_prefills_the_think_block(tmp_path):
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     path = tmp_path / "glm53_no_thinking.py"
     path.write_text(GLM53_NO_THINKING_SOURCE)
diff --git a/tests/test_deepseek_r1.py b/tests/test_deepseek_r1.py
index 41d03e51..d3b992f7 100644
--- a/tests/test_deepseek_r1.py
+++ b/tests/test_deepseek_r1.py
@@ -7,12 +7,16 @@
 
 from functools import lru_cache
 
+import pytest
+
 from renderers import (
     DeepSeekR1Renderer,
     DeepSeekV3Renderer,
     create_renderer,
 )
-from renderers.base import load_tokenizer
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 @lru_cache
diff --git a/tests/test_deepseek_v4.py b/tests/test_deepseek_v4.py
index 4a6b4138..f4207a3f 100644
--- a/tests/test_deepseek_v4.py
+++ b/tests/test_deepseek_v4.py
@@ -6,6 +6,7 @@
 
 import pytest
 from pydantic import TypeAdapter, ValidationError
+
 from renderers import (
     DeepSeekV4Renderer,
     DeepSeekV4RendererConfig,
@@ -13,9 +14,12 @@
     ToolCallParseStatus,
     create_renderer,
 )
-from renderers.base import MODEL_RENDERER_MAP, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP
+from tests.model_assets import load_test_tokenizer as load_tokenizer
 from tests.reference_rendering import render_reference
 
+pytestmark = pytest.mark.model
+
 MODEL = "deepseek-ai/DeepSeek-V4-Flash-0731"
 BOS = "<｜begin▁of▁sentence｜>"
 EOS = "<｜end▁of▁sentence｜>"
diff --git a/tests/test_disabled_thinking_stability.py b/tests/test_disabled_thinking_stability.py
index c6663ecd..dfd8b063 100644
--- a/tests/test_disabled_thinking_stability.py
+++ b/tests/test_disabled_thinking_stability.py
@@ -24,10 +24,14 @@
 
 from functools import lru_cache
 
+import pytest
 from parity import models_for
+
 from renderers import create_renderer
-from renderers.base import load_tokenizer
 from renderers.configs import _config_class_for
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 # One representative model per affected renderer family, each with
 # thinking explicitly disabled.
diff --git a/tests/test_gemma4.py b/tests/test_gemma4.py
index 3360e07f..7d0edf3b 100644
--- a/tests/test_gemma4.py
+++ b/tests/test_gemma4.py
@@ -6,8 +6,12 @@
 from parity import models_for
 
 from renderers import Gemma4Renderer, create_renderer
-from renderers.base import MODEL_RENDERER_MAP, MULTIMODAL_MODELS, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP, MULTIMODAL_MODELS
 from renderers.configs import Gemma4RendererConfig
+from tests.model_assets import load_test_processor
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 _MODELS = tuple(case.model for case in models_for("gemma-checkpoints"))
@@ -298,7 +302,9 @@ def test_parser_recovers_prompt_opened_post_tool_reasoning():
 def _image_renderer():
     """A Gemma 4 renderer with a live ``Gemma4Processor``, or a skip."""
     tokenizer, _ = _gemma4()
-    renderer = Gemma4Renderer(tokenizer)
+    renderer = Gemma4Renderer(
+        tokenizer, processor=load_test_processor("google/gemma-4-31B-it")
+    )
     try:
         renderer._get_processor()
     except (RuntimeError, OSError) as exc:  # pragma: no cover - env dependent
@@ -312,6 +318,7 @@ def _tiny_image():
     return Image.new("RGB", (224, 224), color=(128, 192, 255))
 
 
+@pytest.mark.multimodal
 @pytest.mark.parametrize("size", [(224, 224), (448, 224), (224, 448)])
 def test_real_processor_keeps_one_batched_row_per_image(size):
     """Guard the ``MultiModalFieldConfig.batched('image')`` contract with
@@ -332,6 +339,7 @@ def test_real_processor_keeps_one_batched_row_per_image(size):
     assert placeholder.length > 0
 
 
+@pytest.mark.multimodal
 def test_schema_unified_image_parts_still_expand_image_tokens():
     """``Dataset.from_list`` unifies the Arrow schema across a content list,
     so an image part round-tripped through a dataset carries ``text: None``
@@ -374,6 +382,7 @@ def test_schema_unified_image_parts_still_expand_image_tokens():
     )
 
 
+@pytest.mark.multimodal
 def test_schema_unified_tool_response_image_parts_survive():
     """Same hazard on the tool-response path, which also has to keep
     accepting untyped text parts."""
diff --git a/tests/test_glm53_flash.py b/tests/test_glm53_flash.py
index bf94b4a2..514c7cf3 100644
--- a/tests/test_glm53_flash.py
+++ b/tests/test_glm53_flash.py
@@ -19,6 +19,8 @@
 
 from renderers.client import generate
 
+pytestmark = pytest.mark.model
+
 MODEL = "zai-org/GLM-5.3-Flash-BF16"
 GLM53_MODELS = ["zai-org/GLM-5.3", MODEL]
 
@@ -45,7 +47,7 @@
 @lru_cache(maxsize=None)
 def _load(model=MODEL):
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer(model)
     return tok, create_renderer(tok)
diff --git a/tests/test_glm_tool_name_validation.py b/tests/test_glm_tool_name_validation.py
index 176db262..ed97efdf 100644
--- a/tests/test_glm_tool_name_validation.py
+++ b/tests/test_glm_tool_name_validation.py
@@ -25,9 +25,13 @@
 
 from functools import lru_cache
 
+import pytest
 from parity import models_for
+
 from renderers.base import ToolCallParseStatus
 
+pytestmark = pytest.mark.model
+
 # Both GLM renderers share ``parse_glm`` and are served by the same
 # strict vLLM parser (the ``glm45`` and ``glm47`` aliases both resolve
 # to ``Glm47MoeModelToolParser`` in vLLM ≥ 0.24).
@@ -55,7 +59,7 @@
 @lru_cache(maxsize=None)
 def _load(model: str, renderer_name: str):
     from renderers import config_from_name, create_renderer
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer(model)
     return tok, create_renderer(tok, config_from_name(renderer_name))
diff --git a/tests/test_hy3.py b/tests/test_hy3.py
index db8b0668..8f38abad 100644
--- a/tests/test_hy3.py
+++ b/tests/test_hy3.py
@@ -15,7 +15,10 @@
 import pytest
 
 from renderers import Hy3RendererConfig, create_renderer
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 _MODEL = "tencent/Hy3"
 
diff --git a/tests/test_inkling.py b/tests/test_inkling.py
index 01caaca9..60948e2d 100644
--- a/tests/test_inkling.py
+++ b/tests/test_inkling.py
@@ -22,9 +22,12 @@
 import pytest
 
 from renderers import create_renderer
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
 from renderers.configs import InklingRendererConfig
 from renderers.inkling import InklingRenderer
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 _MODEL = "thinkingmachines/Inkling"
 _SMALL_MODEL = "thinkingmachines/Inkling-Small"
diff --git a/tests/test_laguna_m1.py b/tests/test_laguna_m1.py
index dd9503fc..f91bb876 100644
--- a/tests/test_laguna_m1.py
+++ b/tests/test_laguna_m1.py
@@ -4,6 +4,7 @@
 
 from functools import lru_cache
 
+import pytest
 from pydantic import TypeAdapter
 
 from renderers import (
@@ -12,7 +13,9 @@
     RendererConfig,
     create_renderer,
 )
-from renderers.base import load_tokenizer
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 _MODEL = "poolside/Laguna-M.1"
 
diff --git a/tests/test_laguna_xs21.py b/tests/test_laguna_xs21.py
index add22649..f9dbb130 100644
--- a/tests/test_laguna_xs21.py
+++ b/tests/test_laguna_xs21.py
@@ -22,9 +22,11 @@
 import pytest
 
 from renderers import create_renderer
-from renderers.base import load_tokenizer
 from renderers.configs import LagunaXS21RendererConfig, config_from_name
 from renderers.laguna_xs2 import LagunaXS21Renderer
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 _MODEL = "poolside/Laguna-XS-2.1"
 
diff --git a/tests/test_llama_3.py b/tests/test_llama_3.py
index 86959774..a9950937 100644
--- a/tests/test_llama_3.py
+++ b/tests/test_llama_3.py
@@ -15,11 +15,13 @@
 from renderers import Llama3Renderer, Llama3RendererConfig, create_renderer
 from renderers.base import (
     MODEL_RENDERER_MAP,
-    ParsedResponse,
     TOKENIZER_SOURCE_OVERRIDES,
+    ParsedResponse,
     ToolCallParseStatus,
-    load_tokenizer,
 )
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 # Pinned date for byte-parity tests. Matches the chat template's
 # strftime fallback so we don't have to override on the apply side.
diff --git a/tests/test_load_tokenizer.py b/tests/test_load_tokenizer.py
index aa9855e2..fbfb2430 100644
--- a/tests/test_load_tokenizer.py
+++ b/tests/test_load_tokenizer.py
@@ -12,9 +12,11 @@
 from types import SimpleNamespace
 from unittest.mock import patch
 
+import pytest
+
 from renderers import base
 from renderers.base import TOKENIZER_SOURCE_OVERRIDES, TRUSTED_REVISIONS, load_tokenizer
-
+from tests.model_assets import load_test_tokenizer
 
 # ---------------------------------------------------------------------------
 # Allow-list shape
@@ -50,20 +52,32 @@ def test_trusted_revisions_are_full_shas():
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.parametrize("revision", [None, "a" * 40])
 @patch("transformers.AutoTokenizer.from_pretrained")
-def test_unlisted_model_loads_without_remote_code(mock_from_pretrained):
-    """Default path: trust_remote_code=False, no revision pin."""
-    load_tokenizer("Qwen/Qwen3-0.6B")
+def test_unlisted_model_loads_without_remote_code(mock_from_pretrained, revision):
+    """Ordinary assets can be pinned without enabling remote code."""
+    load_tokenizer("Qwen/Qwen3-0.6B", revision=revision)
     args, kwargs = mock_from_pretrained.call_args
     assert args == ("Qwen/Qwen3-0.6B",)
-    assert kwargs == {"trust_remote_code": False}
+    expected = {"trust_remote_code": False}
+    if revision is not None:
+        expected["revision"] = revision
+    assert kwargs == expected
 
 
+@pytest.mark.parametrize(
+    "revision", [None, TRUSTED_REVISIONS["moonshotai/Kimi-K2.5"], "a" * 40]
+)
 @patch("transformers.AutoTokenizer.from_pretrained")
-def test_kimi_loads_with_pinned_revision(mock_from_pretrained):
+def test_kimi_loads_with_pinned_revision(mock_from_pretrained, revision):
     """Kimi-K2 family: trust_remote_code=True, revision pinned to the
     sha listed in TRUSTED_REVISIONS — never a branch name."""
-    load_tokenizer("moonshotai/Kimi-K2.5")
+    if revision is not None and revision != TRUSTED_REVISIONS["moonshotai/Kimi-K2.5"]:
+        with pytest.raises(ValueError, match="reviewed remote-code revision"):
+            load_tokenizer("moonshotai/Kimi-K2.5", revision=revision)
+        mock_from_pretrained.assert_not_called()
+        return
+    load_tokenizer("moonshotai/Kimi-K2.5", revision=revision)
     args, kwargs = mock_from_pretrained.call_args
     assert args == ("moonshotai/Kimi-K2.5",)
     assert kwargs == {
@@ -150,22 +164,24 @@ def __call__(self, *args, **kwargs):
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.model
 def test_load_tokenizer_real_qwen_works_without_remote_code():
     """End-to-end: an unlisted model loads successfully without
     trust_remote_code. Qwen tokenizers don't ship custom Python."""
-    tok = load_tokenizer("Qwen/Qwen3-0.6B")
+    tok = load_test_tokenizer("Qwen/Qwen3-0.6B")
     assert tok is not None
     # Smoke: the tokenizer can encode a basic string.
     ids = tok.encode("hello", add_special_tokens=False)
     assert len(ids) > 0
 
 
+@pytest.mark.model
 def test_load_tokenizer_real_kimi_uses_pinned_revision():
     """End-to-end: Kimi-K2.5 loads via the pinned-revision path. The
     parity tests already exercise this path — this test pins the
     contract that ``load_tokenizer`` is the only sanctioned entry
     point for the trusted-revision allow-list."""
-    tok = load_tokenizer("moonshotai/Kimi-K2.5")
+    tok = load_test_tokenizer("moonshotai/Kimi-K2.5")
     assert tok is not None
     ids = tok.encode("hello", add_special_tokens=False)
     assert len(ids) > 0
diff --git a/tests/test_message_indices.py b/tests/test_message_indices.py
index d6471e37..8f38217c 100644
--- a/tests/test_message_indices.py
+++ b/tests/test_message_indices.py
@@ -43,8 +43,8 @@
 import pytest
 
 from renderers import build_training_sample
-from renderers.glm45 import GLM45Renderer
 from renderers.glm5 import GLM5Renderer
+from renderers.glm45 import GLM45Renderer
 from tests.reference_rendering import render_reference
 
 
@@ -140,6 +140,7 @@ def test_message_indices_in_range(model_name, renderer):
     )
 
 
+@pytest.mark.model
 def test_kimi_k2_unknown_role_message_indices():
     """Regression for the ``i`` vs ``oi`` mistake in
     ``KimiK2Renderer.render``'s unknown-role fallback. To surface, we
@@ -159,7 +160,7 @@ def test_kimi_k2_unknown_role_message_indices():
     caller list of length 2.
     """
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer("moonshotai/Kimi-K2-Instruct")
     renderer = create_renderer(tok)
diff --git a/tests/test_multimodal.py b/tests/test_multimodal.py
index aadcf152..3d8aad79 100644
--- a/tests/test_multimodal.py
+++ b/tests/test_multimodal.py
@@ -15,7 +15,7 @@
    the combined message list.
 
 Tests skip per-pair when:
-- The HF snapshot isn't cached locally (network-free CI mode).
+- Optional private model metadata is unavailable locally.
 - The model lists a modality the renderer doesn't yet support
   (``NotImplementedError`` in ``render``).
 - ``Pillow`` / ``torch`` are missing.
@@ -23,9 +23,6 @@
 
 from __future__ import annotations
 
-import os
-from pathlib import Path
-
 import pytest
 
 from renderers import (
@@ -33,8 +30,10 @@
     Qwen3VLRenderer,
     create_renderer,
 )
-from renderers.base import MODEL_RENDERER_MAP, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP
 from renderers.configs import _config_class_for
+from tests.model_assets import load_test_processor, model_revision
+from tests.model_assets import load_test_tokenizer as load_tokenizer
 
 
 def _config_for_model(model_name: str, **kwargs):
@@ -55,28 +54,25 @@ def _config_with_add_vision_id(model_name: str, add_vision_id: bool):
 import numpy as np  # noqa: E402
 from PIL import Image  # noqa: E402
 
+pytestmark = [pytest.mark.model, pytest.mark.multimodal]
+
 
 # ---------------------------------------------------------------------------
-# Local-snapshot gating — skip when the HF cache doesn't have the model.
+# Optional private metadata stays cache-only; public assets are pinned.
 # ---------------------------------------------------------------------------
 
 
-def _hf_snapshot_cached(model_name: str) -> bool:
-    """True iff the HF hub cache has at least one snapshot for ``model_name``.
+def _model_assets_available(model_name: str) -> bool:
+    """Download pinned public assets; keep optional private metadata cache-only."""
+    if model_name != "nvidia/NVIDIA-Nemotron-3.5-Super-EA-09112026":
+        model_revision(model_name)
+        return True
 
-    Avoids pulling weights / configs over the network during test runs.
-    Mirrors the convention used elsewhere in this repo (test_qwen35_size_coverage)
-    of relying on the user having pre-fetched relevant models.
-    """
-    cache = (
-        Path(os.environ.get("HF_HOME") or Path.home() / ".cache" / "huggingface")
-        / "hub"
-    )
-    safe = "models--" + model_name.replace("/", "--")
-    snapshots = cache / safe / "snapshots"
-    if not snapshots.is_dir():
+    try:
+        model_revision(model_name)
+    except KeyError:
         return False
-    return any(p.is_dir() for p in snapshots.iterdir())
+    return True
 
 
 # ---------------------------------------------------------------------------
@@ -104,36 +100,12 @@ def _modality_cases():
 _loaded: dict[str, tuple] = {}
 
 
-# Models whose processors need ``trust_remote_code=True`` (custom Python
-# in the repo) and still use a pinned revision. Nemotron 3.5 is handled
-# separately below so its selected model revision is not overridden.
-_PROCESSOR_TRUSTED_REVISIONS: dict[str, str] = {
-    "moonshotai/Kimi-K2.5": "4d01dfe0332d63057c186e0b262165819efb6611",
-    "moonshotai/Kimi-K2.6": "2755962d07cb42aa2d988a35bcb65cd4a9c2de82",
-}
-
-
 def _load_processor_and_renderer(model_name: str):
     if model_name not in _loaded:
-        from transformers import AutoProcessor
-
         tokenizer = load_tokenizer(model_name)
-        revision = _PROCESSOR_TRUSTED_REVISIONS.get(model_name)
-        if model_name.startswith("nvidia/NVIDIA-Nemotron-3.5"):
-            processor = AutoProcessor.from_pretrained(
-                model_name,
-                trust_remote_code=True,
-            )
-        elif revision is not None:
-            processor = AutoProcessor.from_pretrained(
-                model_name,
-                trust_remote_code=True,
-                revision=revision,
-            )
-        else:
-            processor = AutoProcessor.from_pretrained(model_name)
+        processor = load_test_processor(model_name)
         renderer = create_renderer(tokenizer)
-        # Inject processor so the renderer doesn't try to fetch it lazily.
+        # Inject the pinned processor instead of lazily resolving the model again.
         if hasattr(renderer, "_processor") and renderer._processor is None:
             renderer._processor = processor
         _loaded[model_name] = (tokenizer, processor, renderer)
@@ -626,8 +598,10 @@ def test_multimodal_byte_parity_vs_processor(mm_model_name, modality, tiny_image
     token ids regardless of whether the prompt is templated server-side
     (MITO) or rendered client-side via this package.
     """
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, processor, renderer = _load_processor_and_renderer(mm_model_name)
@@ -659,8 +633,10 @@ def test_multimodal_byte_parity_vs_processor(mm_model_name, modality, tiny_image
 )
 def test_multimodal_placeholders_match_pad_runs(mm_model_name, modality, tiny_image):
     """``mm_placeholders`` exactly cover the runs of the modality's pad token."""
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, _, renderer = _load_processor_and_renderer(mm_model_name)
@@ -723,8 +699,10 @@ def test_multimodal_bridge_extends_and_carries_mm_data(
        the new turn's image, plus its placeholder is recorded with an
        absolute offset inside the bridged sequence.
     """
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, processor, _ = _load_processor_and_renderer(mm_model_name)
@@ -849,8 +827,8 @@ def test_inkling_bridge_does_not_mutate_prior_mm_data(tiny_image):
     the merge copies the per-modality lists, not just the outer dict, so a
     carried-forward image doesn't grow the prior turn's list in place."""
     model = "thinkingmachines/Inkling"
-    if not _hf_snapshot_cached(model):
-        pytest.skip(f"{model}: HF snapshot not cached locally")
+    if not _model_assets_available(model):
+        pytest.skip(f"{model}: optional private model metadata not cached locally")
 
     _, processor, renderer = _load_processor_and_renderer(model)
     initial = [
@@ -899,7 +877,7 @@ def test_modality_registry_models_route_to_renderer():
     via ``create_renderer(renderer='auto')``. Guards against drift between
     the multimodal registry and ``MODEL_RENDERER_MAP``."""
     for model_name in MULTIMODAL_MODELS:
-        if not _hf_snapshot_cached(model_name):
+        if not _model_assets_available(model_name):
             continue
         tokenizer = load_tokenizer(model_name)
         renderer = create_renderer(tokenizer)
@@ -925,8 +903,10 @@ def test_tool_response_image_byte_parity(mm_model_name, modality, tiny_image):
     """
     if modality != "image":
         pytest.skip("Tool-response media path is image-only for now.")
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, processor, renderer = _load_processor_and_renderer(mm_model_name)
@@ -1011,8 +991,10 @@ def test_add_vision_id_parity_vs_processor(
     ``processor.apply_chat_template(messages, add_vision_id=<value>)``
     token-for-token after image expansion.
     """
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, processor, _ = _load_processor_and_renderer(mm_model_name)
@@ -1062,8 +1044,10 @@ def test_bridge_refuses_when_add_vision_id_loses_prior_count(
     ``apply_chat_template`` and a full re-render. The bridge must
     refuse (return None) so the caller falls back to a full re-render.
     """
-    if not _hf_snapshot_cached(mm_model_name):
-        pytest.skip(f"{mm_model_name}: HF snapshot not cached locally")
+    if not _model_assets_available(mm_model_name):
+        pytest.skip(
+            f"{mm_model_name}: optional private model metadata not cached locally"
+        )
 
     kit = _modality_kit(modality, mm_model_name)
     tokenizer, processor, _ = _load_processor_and_renderer(mm_model_name)
@@ -1137,8 +1121,8 @@ def test_qwen3_vl_renderer_exposes_image_modality():
     registry → load → create_renderer(auto) → expected concrete class.
     """
     model = "Qwen/Qwen3-VL-4B-Instruct"
-    if not _hf_snapshot_cached(model):
-        pytest.skip(f"{model}: HF snapshot not cached locally")
+    if not _model_assets_available(model):
+        pytest.skip(f"{model}: optional private model metadata not cached locally")
     tokenizer = load_tokenizer(model)
     renderer = create_renderer(tokenizer)
     assert isinstance(renderer, Qwen3VLRenderer)
diff --git a/tests/test_muse_glimmer.py b/tests/test_muse_glimmer.py
index 500898ee..41dc0172 100644
--- a/tests/test_muse_glimmer.py
+++ b/tests/test_muse_glimmer.py
@@ -17,8 +17,11 @@
 from functools import lru_cache
 
 import pytest
+
 from renderers.base import ToolCallParseStatus
 
+pytestmark = pytest.mark.model
+
 MODEL = "meta-models/Muse-Glimmer-30B"
 DATE = "2026-10-16"
 IMAGE = {"type": "image_url", "image_url": {"url": "https://example.com/a.png"}}
@@ -70,8 +73,8 @@ def _invoke(name: str, *params: tuple[str, str]) -> str:
 @lru_cache(maxsize=None)
 def _load(**config):
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
     from renderers.configs import MuseGlimmerRendererConfig
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer(MODEL)
     return tok, create_renderer(
@@ -132,7 +135,7 @@ def _token(text: str) -> int:
 
 def test_create_renderer_resolves_muse_glimmer():
     from renderers import create_renderer
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     renderer = create_renderer(load_tokenizer(MODEL))
     assert type(renderer).__name__ == "MuseGlimmerRenderer"
diff --git a/tests/test_nemotron35_vlm.py b/tests/test_nemotron35_vlm.py
index 76a86cab..f1bfd28d 100644
--- a/tests/test_nemotron35_vlm.py
+++ b/tests/test_nemotron35_vlm.py
@@ -12,7 +12,6 @@
 from renderers.base import MODEL_RENDERER_MAP, MULTIMODAL_MODELS, PlaceholderRange
 from renderers.nemotron3 import Nemotron35Renderer
 
-
 MODEL = "nvidia/NVIDIA-Nemotron-3.5-Super-EA-09112026"
 
 
@@ -277,6 +276,8 @@ def test_nemotron35_loads_approved_remote_processor_without_revision(monkeypatch
     ]
 
 
+@pytest.mark.model
+@pytest.mark.multimodal
 def test_real_processor_token_parity_when_metadata_is_available():
     metadata = os.environ.get("NEMOTRON35_VLM_METADATA")
     if metadata is None:
diff --git a/tests/test_offsetless_tokenizers.py b/tests/test_offsetless_tokenizers.py
index 255c2c68..85174c8b 100644
--- a/tests/test_offsetless_tokenizers.py
+++ b/tests/test_offsetless_tokenizers.py
@@ -7,7 +7,7 @@
 import pytest
 
 from renderers import create_renderer
-from renderers.base import load_tokenizer
+from tests.model_assets import load_test_tokenizer as load_tokenizer
 
 
 class OffsetlessTokenizer:
@@ -181,6 +181,7 @@ def test_offsetless_contract_across_renderer_bridges(
         _assert_offsetless_contract(expected, actual, renderer)
 
 
+@pytest.mark.model
 def test_hy3_offsetless_preserves_multiple_system_message_indices() -> None:
     """Hy3's joined system header must not erase either caller message."""
 
@@ -199,6 +200,7 @@ def test_hy3_offsetless_preserves_multiple_system_message_indices() -> None:
     assert {0, 1}.issubset(actual.message_indices)
 
 
+@pytest.mark.model
 @pytest.mark.parametrize(
     "deepseek_model",
     ["deepseek-ai/DeepSeek-V3", "deepseek-ai/DeepSeek-R1"],
diff --git a/tests/test_parity.py b/tests/test_parity.py
index d4711c68..928934bd 100644
--- a/tests/test_parity.py
+++ b/tests/test_parity.py
@@ -16,7 +16,6 @@
 from typing import Annotated, Any, Literal, Mapping, Union, cast, get_args, get_origin
 
 import pytest
-
 from parity import (
     KWARG_VALUES,
     MODEL_CATALOG,
@@ -26,9 +25,11 @@
     kwarg_combinations,
     scenario_is_valid,
 )
+
 from renderers import create_renderer
-from renderers.base import MODEL_RENDERER_MAP, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP
 from renderers.configs import RendererConfig, _config_class_for
+from tests.model_assets import load_test_tokenizer as load_tokenizer
 from tests.reference_rendering import (
     DEFAULT_REFERENCE_ORACLE,
     REFERENCE_ORACLES,
@@ -38,6 +39,8 @@
     render_reference,
 )
 
+pytestmark = pytest.mark.model
+
 
 def _id(case: ModelCase, scenario: Scenario, kwargs: Mapping[str, Any]) -> str:
     values = ",".join(f"{key}={value!r}" for key, value in kwargs.items())
diff --git a/tests/test_parse_response.py b/tests/test_parse_response.py
index be5afbd7..ee9a70ea 100644
--- a/tests/test_parse_response.py
+++ b/tests/test_parse_response.py
@@ -7,8 +7,12 @@
 from functools import lru_cache
 
 import pytest
+
 from renderers import create_renderer
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 @lru_cache
diff --git a/tests/test_parsers.py b/tests/test_parsers.py
index a81a5eb3..00b34b37 100644
--- a/tests/test_parsers.py
+++ b/tests/test_parsers.py
@@ -6,7 +6,7 @@
 
 import pytest
 
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
 from renderers.parsers import (
     REASONING_PARSERS,
     TOOL_PARSERS,
@@ -15,6 +15,9 @@
     get_reasoning_parser,
     get_tool_parser,
 )
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 def test_registries_nonempty():
diff --git a/tests/test_qwen35_size_coverage.py b/tests/test_qwen35_size_coverage.py
index b609d14f..76abdc7d 100644
--- a/tests/test_qwen35_size_coverage.py
+++ b/tests/test_qwen35_size_coverage.py
@@ -18,8 +18,8 @@
 from parity import MODEL_CATALOG
 
 from renderers import Qwen35Renderer, Qwen35RendererConfig, create_renderer
-from renderers.base import MODEL_RENDERER_MAP, load_tokenizer
-
+from renderers.base import MODEL_RENDERER_MAP
+from tests.model_assets import load_test_tokenizer as load_tokenizer
 
 _QWEN35_IN_MAP = {
     case.model for case in MODEL_CATALOG if case.resolved_renderer == "qwen3.5"
@@ -53,6 +53,7 @@ def test_no_other_qwen35_sizes_silently_added():
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.model
 @pytest.mark.parametrize(
     "qwen35_model,expected_default",
     [
diff --git a/tests/test_qwen38.py b/tests/test_qwen38.py
index 9fbcacf1..6ccb4e25 100644
--- a/tests/test_qwen38.py
+++ b/tests/test_qwen38.py
@@ -2,8 +2,8 @@
 
 from __future__ import annotations
 
-from functools import lru_cache
 from copy import deepcopy
+from functools import lru_cache
 
 import pytest
 from pydantic import TypeAdapter
@@ -14,7 +14,10 @@
     RendererConfig,
     create_renderer,
 )
-from renderers.base import MODEL_RENDERER_MAP, MULTIMODAL_MODELS, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP, MULTIMODAL_MODELS
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 TOOLS = [
diff --git a/tests/test_reasoning_boundaries.py b/tests/test_reasoning_boundaries.py
index eecf0281..1cade4cf 100644
--- a/tests/test_reasoning_boundaries.py
+++ b/tests/test_reasoning_boundaries.py
@@ -4,9 +4,13 @@
 
 import pytest
 from parity import MODEL_CATALOG
+
 from renderers import create_renderer
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
 from renderers.configs import config_from_name
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 # Exercise every registered renderer, including inherited family variants.
 _CASES = {}
diff --git a/tests/test_roundtrip.py b/tests/test_roundtrip.py
index c1759b84..56675072 100644
--- a/tests/test_roundtrip.py
+++ b/tests/test_roundtrip.py
@@ -14,9 +14,12 @@
 from replay import assert_evidence, replay_client
 
 from renderers import create_renderer
-from renderers.base import MODEL_RENDERER_MAP, ToolCallParseStatus, load_tokenizer
+from renderers.base import MODEL_RENDERER_MAP, ToolCallParseStatus
 from renderers.client import generate
 from renderers.configs import config_from_name
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 # Both matrices visit checkpoints in groups; reuse within each group only.
diff --git a/tests/test_tagged_rollout.py b/tests/test_tagged_rollout.py
index 68677a71..c34170e6 100644
--- a/tests/test_tagged_rollout.py
+++ b/tests/test_tagged_rollout.py
@@ -11,9 +11,12 @@
 from replay import assert_evidence, replay_client
 
 from renderers import create_renderer
-from renderers.base import ToolCallParseStatus, load_tokenizer
+from renderers.base import ToolCallParseStatus
 from renderers.client import generate
 from renderers.configs import config_from_name
+from tests.model_assets import load_test_tokenizer as load_tokenizer
+
+pytestmark = pytest.mark.model
 
 
 @lru_cache(maxsize=1)
diff --git a/tests/test_tool_arg_type_preservation.py b/tests/test_tool_arg_type_preservation.py
index ec12a3e2..d8196417 100644
--- a/tests/test_tool_arg_type_preservation.py
+++ b/tests/test_tool_arg_type_preservation.py
@@ -30,6 +30,8 @@
 
 from renderers.muse_glimmer import MuseGlimmerRenderer
 
+pytestmark = pytest.mark.model
+
 
 # (HuggingFace model name, renderer name). Two JSON-shaped controls
 # (string types already preserved by the wire format) + five XML-style
@@ -40,7 +42,7 @@
 @lru_cache(maxsize=None)
 def _load(model: str, renderer_name: str):
     from renderers import config_from_name, create_renderer
-    from renderers.base import load_tokenizer
+    from tests.model_assets import load_test_tokenizer as load_tokenizer
 
     tok = load_tokenizer(model)
     return tok, create_renderer(tok, config_from_name(renderer_name))