diff --git a/CHANGELOG.md b/CHANGELOG.md
index ed5131f0..7b012d15 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 ## [Unreleased]
 
 ### Added
+- Added `PerFileConverterMixin`, which converts each source file on its own and merges the parts; ES uses it, as its 50 provinces do not fit in memory together.
 - Added `pixi run check-hcat`, which compares the HCAT mapping tables the converters read online with the taxonomy.
 - Added `FiboaDuckDBBaseConverter` for SQL-based conversion of large Parquet sources.
 - Added `PerFileBaseConverter` to process multi-file sources incrementally.
diff --git a/fiboa_cli/conversion/per_file.py b/fiboa_cli/conversion/per_file.py
new file mode 100644
index 00000000..d0161f19
--- /dev/null
+++ b/fiboa_cli/conversion/per_file.py
@@ -0,0 +1,34 @@
+import os
+from tempfile import TemporaryDirectory
+
+from vecorel_cli.conversion.duckdb import DuckDBBaseConverter
+
+
+class PerFileConverterMixin:
+    """Converts each source file on its own and merges the parts, for sources too large to read at once."""
+
+    def convert(self, output_file, cache=None, input_files=None, variant=None, **kwargs) -> str:
+        self.select_variant(variant)
+        urls = input_files or self.get_urls()
+        if not urls or len(urls) == 1:
+            return super().convert(
+                output_file, cache=cache, input_files=input_files, variant=variant, **kwargs
+            )
+
+        directory = os.path.dirname(output_file) or "."
+        os.makedirs(directory, exist_ok=True)
+        with TemporaryDirectory(dir=directory) as parts_folder:
+            parts = []
+            for index, (uri, target) in enumerate(urls.items()):
+                self.info(f"Converting source {index + 1}/{len(urls)}: {uri}")
+                part = os.path.join(parts_folder, f"part_{index}.parquet")
+                super().convert(
+                    part, cache=cache, input_files={uri: target}, variant=variant, **kwargs
+                )
+                parts.append(part)
+
+            self.info(f"Merging {len(parts)} parts into {output_file}")
+            merge_options = ("compression", "compression_level", "geoparquet_version")
+            return DuckDBBaseConverter().merge_parquet(
+                parts, output_file, **{k: kwargs[k] for k in merge_options if k in kwargs}
+            )
diff --git a/fiboa_cli/datasets/es.py b/fiboa_cli/datasets/es.py
index 742c803e..d88da6fe 100644
--- a/fiboa_cli/datasets/es.py
+++ b/fiboa_cli/datasets/es.py
@@ -4,10 +4,12 @@
 from vecorel_cli.vecorel.extensions import ADMIN_DIVISION
 
 from ..conversion.fiboa_converter import FiboaBaseConverter
+from ..conversion.per_file import PerFileConverterMixin
 from .commons.hcat import AddHCATMixin
 
 
-class Converter(AddHCATMixin, FiboaBaseConverter):
+# 50 provincial GeoPackages that do not fit in memory together
+class Converter(PerFileConverterMixin, AddHCATMixin, FiboaBaseConverter):
     id = "es"
     short_name = "Spain"
     title = "Spain Declared Crops (Cultivos Declarados SIGPAC)"
diff --git a/tests/test_per_file.py b/tests/test_per_file.py
new file mode 100644
index 00000000..e30f91d5
--- /dev/null
+++ b/tests/test_per_file.py
@@ -0,0 +1,24 @@
+import shutil
+
+import geopandas as gpd
+
+from fiboa_cli.converters import Converters
+
+FIXTURE = "tests/data-files/convert/es/1501_ALAVA_cd_2025_20250105.gpkg.zip"
+
+
+def test_each_source_is_converted_on_its_own_and_merged(tmp_path):
+    from fiboa_cli import Registry  # noqa
+
+    copy = tmp_path / "copy.gpkg.zip"
+    shutil.copy(FIXTURE, copy)
+    output = tmp_path / "es.parquet"
+
+    Converters().load("es").convert(
+        str(output), input_files={FIXTURE: ["*.gpkg"], str(copy): ["*.gpkg"]}
+    )
+
+    merged = gpd.read_parquet(output)
+    assert len(merged) == 20  # 10 rows per source
+    assert merged["id"].is_unique
+    assert set(tmp_path.iterdir()) == {copy, output}  # the parts are gone