diff --git a/fanan/portal/__init__.py b/fanan/portal/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/fanan/portal/huggingface.py b/fanan/portal/huggingface.py
new file mode 100644
index 0000000..da38cf9
--- /dev/null
+++ b/fanan/portal/huggingface.py
@@ -0,0 +1,66 @@
+import os
+import tempfile
+import urllib
+import urllib.parse
+from dataclasses import dataclass
+
+import fsspec
+from transformers import (
+    AutoTokenizer,
+    PreTrainedTokenizer,
+    PreTrainedTokenizerFast,
+)
+
+PYTORCH_MODEL = "pytorch_model.bin"
+SAFE_TENSORS_MODEL = "model.safetensors"
+
+
+@dataclass(frozen=True)
+class RepoRef:
+    model_name_or_path: str
+    revision: str | None = None
+
+    @staticmethod
+    def from_string(name: str) -> "RepoRef":
+        if "@" not in name:
+            return RepoRef(name)
+        model_name_or_path, revision = name.split("@")
+        return RepoRef(model_name_or_path, revision)
+
+    def __str__(self) -> str:
+        if self.revision is None:
+            return self.model_name_or_path
+        return f"{self.model_name_or_path}@{self.revision}"
+
+    def __repr__(self) -> str:
+        return f"RemoteRev({self.model_name_or_path!r}, {self.revision!r})"
+
+
+def _is_url_like(path: str) -> bool:
+    return urllib.parse.urlparse(path).scheme != ""
+
+
+def load_tokenizer(
+    model_name_or_path: str,
+    revision: str | None = None,
+    local_cache_dir: str | None = None,
+    trust_remote_code: bool = True,
+) -> PreTrainedTokenizer | PreTrainedTokenizerFast:
+    """Load a tokenizer from a model name or path, with optional revision and local cache directory."""
+    if _is_url_like(model_name_or_path):
+        if revision is not None:
+            raise ValueError("revision is not supported for URLs")
+
+        if local_cache_dir is None:
+            local_cache_dir = tempfile.mkdtemp()
+
+        fs, path = fsspec.core.url_to_fs(model_name_or_path)
+        fs.get(path, local_cache_dir, recursive=True)
+        base_path = os.path.basename(path)
+
+        return AutoTokenizer.from_pretrained(
+            os.path.join(local_cache_dir, base_path),
+            trust_remote_code=trust_remote_code,
+        )
+    else:
+        return AutoTokenizer.from_pretrained(model_name_or_path, revision=revision, trust_remote_code=trust_remote_code)