@@ -1303,6 +1303,26 @@ def greeting_pipeline() -> Out[str]:
 
 Use `unwrap=["items", "metadata"]` to unwrap multiple dict parameters. The caller owns the key names; keys may contain letters, numbers, `_`, and `-`, and become `param__<key>` component inputs. Empty dicts are rejected because they do not define a component interface. If a generated name would collide with a fixed parameter or another generated name, compile fails before writing artifacts. Equivalent key sets are canonicalized for schema hashing, so two call sites with the same keys in different insertion orders dedupe to the same component fragment.
 
+Unwrapped values can also be artifacts. Annotate the dict values with `components.InputPath(...)` (use the builtin `dict[...]`; `typing.Dict` rejects non-type arguments) and every key becomes an `inputPath` input of that type. At runtime the function receives a dict mapping each key to the local path of its downloaded artifact:
+
+```python
+@task(image="python:3.12", unwrap="tables")
+def merge_parquets(
+    tables: dict[str, components.InputPath("ApacheParquet")],
+    merged_path: components.OutputPath("ApacheParquet"),
+):
+    import pyarrow as pa
+    import pyarrow.parquet as pq
+
+    pq.write_table(pa.concat_tables([pq.read_table(tables[key]) for key in sorted(tables)]), merged_path)
+
+
+# In a pipeline: any number of upstream Parquet outputs, keyed by the caller.
+merged = merge_parquets.named("merge")(tables={"shop": shop.table, "catalog": catalog.table})
+```
+
+Artifact unwraps persist `"kind": "input_path"` on each key (and `"value_kind": "input_path"` on the parameter) in `local_from_python.unwrapped_inputs`, so hydrate regenerates the same `inputPath` interface. Value unwraps omit these fields; entries without them are value inputs. `OutputPath` dict values are rejected.
+
 #### Pipeline run submission and validation
 
 `submit` hydrates refs by default and builds an API submit payload with `root_task.componentRef.spec`. Use `--no-hydrate` to submit the local YAML structure as-is. Use `--dry-run` to print the payload without creating a run.
 
@@ -14,6 +14,6 @@
 try:
     __version__ = metadata_version("tangle-cli")
 except PackageNotFoundError:
-    __version__ = "0.1.32"
+    __version__ = "0.1.33"
 
 __all__ = ["TangleDynamicDiscoveryClient", "__version__"]
@@ -345,30 +345,41 @@ def _dict_value_annotation(annotation: Any) -> Any:
     return None
 
 
-def _unwrapped_value_type(annotation: Any) -> tuple[str, str]:
-    """Infer the Tangle scalar type/deserializer for unwrapped dict values.
+_UNWRAPPED_KINDS = ("input", "input_path")
+
+
+def _unwrapped_value_spec(annotation: Any) -> tuple[str, str, Literal["input", "input_path"]]:
+    """Infer the Tangle type, deserializer and kind for unwrapped dict values.
 
     Args:
         annotation: The resolved annotation for the original dict parameter.
 
     Returns:
-        ``(tangle_type, deserializer)`` for each synthetic flattened input.
-        Unknown ``Any`` values default to ``String``/``str`` to match existing
+        ``(tangle_type, deserializer, kind)`` for each synthetic flattened
+        input. ``kind`` is ``"input"`` for value inputs (``inputValue``
+        placeholders) and ``"input_path"`` when the dict values are annotated
+        as ``InputPath(...)`` artifacts (``inputPath`` placeholders; the
+        runtime receives one local file path per key). Unknown ``Any`` values
+        default to ``String``/``str`` value inputs to match existing
         unannotated input behavior.
 
     Raises:
         ValueError: If the unwrap target is not annotated as ``dict`` or
-            ``dict[..., ...]``.
+            ``dict[..., ...]``, or its values are ``OutputPath`` annotations.
     """
     value_annotation = _dict_value_annotation(annotation)
     if value_annotation is None:
         raise ValueError("@task unwrap parameters must be annotated as dict[...] or dict")
     if value_annotation is Any:
-        return "String", "str"
+        return "String", "str", "input"
     tangle_type, deserializer, kind = _resolve_annotation(value_annotation)
-    if kind != "input" or not tangle_type:
-        return "String", "str"
-    return tangle_type, deserializer
+    if kind == "input_path":
+        return tangle_type or "String", "str", "input_path"
+    if kind == "output":
+        raise ValueError("@task unwrap parameters cannot have OutputPath values; only inputs can be unwrapped")
+    if not tangle_type:
+        return "String", "str", "input"
+    return tangle_type, deserializer, "input"
 
 
 def _normalize_unwrapped_input_keys(
@@ -435,7 +446,11 @@ def build_unwrapped_inputs_schema(
         writes this into the resolver sidecar so hydrate can regenerate the
         identical component interface without access to the original call-site
         dict. Each key entry records the original key, synthetic input name,
-        inferred Tangle type, and optional flag.
+        inferred Tangle type, and optional flag. Artifact unwraps
+        (``dict[str, InputPath(...)]``) additionally record
+        ``"kind": "input_path"`` on each key and ``"value_kind":
+        "input_path"`` on the parameter entry; value unwraps omit both keys so
+        their persisted schema (and fragment hash) is unchanged.
 
     Raises:
         ValueError: If an unwrap parameter is absent from the function
@@ -464,7 +479,7 @@ def build_unwrapped_inputs_schema(
         if param_name not in signature.parameters:
             raise ValueError(f"@task unwrap parameter {param_name!r} is not in the function signature")
         annotation = resolved_hints.get(param_name, signature.parameters[param_name].annotation)
-        tangle_type, _deserializer = _unwrapped_value_type(annotation)
+        tangle_type, _deserializer, kind = _unwrapped_value_spec(annotation)
         input_prefix = f"{param_name}__"
         key_specs = []
         for key in keys:
@@ -480,19 +495,23 @@ def build_unwrapped_inputs_schema(
                     f"which collides with unwrap parameter {seen_synthetic_names[input_name]!r}"
                 )
             seen_synthetic_names[input_name] = param_name
-            key_specs.append(
-                {
-                    "key": key,
-                    "input_name": input_name,
-                    "type": tangle_type,
-                    "optional": False,
-                }
-            )
-        schema[param_name] = {
+            key_spec: dict[str, Any] = {
+                "key": key,
+                "input_name": input_name,
+                "type": tangle_type,
+                "optional": False,
+            }
+            if kind != "input":
+                key_spec["kind"] = kind
+            key_specs.append(key_spec)
+        param_schema: dict[str, Any] = {
             "input_prefix": input_prefix,
             "value_type": tangle_type,
-            "keys": key_specs,
         }
+        if kind != "input":
+            param_schema["value_kind"] = kind
+        param_schema["keys"] = key_specs
+        schema[param_name] = param_schema
     return schema
 
 
@@ -517,28 +536,40 @@ def _expand_unwrapped_param(
         A ``ParamInfo`` for each flattened key, using names like
         ``run_data__shop``. The returned params retain ``source_param`` and
         ``source_key`` so generated runtime code can re-wrap CLI arguments into
-        the original dict before invoking the function.
+        the original dict before invoking the function. The persisted
+        ``kind``/``value_kind`` (default ``"input"``) decides whether each
+        key becomes an ``inputValue`` or an ``inputPath`` input; artifact
+        inputs always receive the local path as a plain ``str``.
 
     Raises:
-        ValueError: If ``annotation`` is not dict-like.
+        ValueError: If ``annotation`` is not dict-like, or the schema records
+            an unknown kind.
     """
-    _unwrapped_value_type(annotation)  # validates dict-like annotation
+    _unwrapped_value_spec(annotation)  # validates dict-like annotation
+    default_kind = schema.get("value_kind") or "input"
     expanded: list[ParamInfo] = []
     for key_spec in schema.get("keys", []) or []:
         key = str(key_spec["key"])
         input_name = str(key_spec.get("input_name") or f"{param.name}__{key}")
         tangle_type = str(key_spec.get("type") or schema.get("value_type") or "String")
+        kind = key_spec.get("kind") or default_kind
+        if kind not in _UNWRAPPED_KINDS:
+            raise ValueError(
+                f"@task unwrap parameter {param.name!r} key {key!r} has unsupported kind {kind!r}; "
+                f"expected one of {list(_UNWRAPPED_KINDS)}"
+            )
+        deserializer = "str" if kind == "input_path" else _TYPE_TO_DESERIALIZER.get(tangle_type, "str")
         expanded.append(
             ParamInfo(
                 name=input_name,
                 yaml_name=input_name,
                 python_type=str(annotation) if annotation is not inspect.Parameter.empty else None,
                 tangle_type=tangle_type,
-                kind="input",
+                kind=kind,
                 description=doc_dict.get(param.name),
                 default=inspect.Parameter.empty,
                 optional=bool(key_spec.get("optional", False)),
-                deserializer=_TYPE_TO_DESERIALIZER.get(tangle_type, "str"),
+                deserializer=deserializer,
                 source_param=param.name,
                 source_key=key,
             )
 
@@ -123,7 +123,10 @@ def task(
             For example ``unwrap="run_data"`` turns
             ``run_data={"run_id_1": task.output}`` into a component input
             named ``run_data__run_id_1`` and reconstructs the original dict in
-            the generated runtime wrapper.
+            the generated runtime wrapper. Values annotated as
+            ``dict[str, components.InputPath(T)]`` become ``inputPath``
+            artifact inputs, and the reconstructed dict maps each key to the
+            artifact's local path.
 
     Returns:
         A decorator that, given the user's function, returns a
 
@@ -1,6 +1,6 @@
 [project]
 name = "tangle-cli"
-version = "0.1.32"
+version = "0.1.33"
 description = "CLI for Tangle, the open-source ML pipeline orchestration platform"
 readme = "README.md"
 authors = [