Limits

Readers and writers enforce these ceilings before allocating or decompressing content. They are safety maxima: you can lower them for a call, never raise them.

ArchiveLimits

Field	Default and maximum
max_members	6 (minimum 2)
max_archive_bytes	512 MiB
max_compressed_bytes	512 MiB
max_member_bytes	512 MiB
max_total_bytes	1 GiB, all members uncompressed
max_compression_ratio	200
max_manifest_bytes	8 MiB
max_structure_bytes	256 MiB
max_schema_bytes	16 MiB
max_samples_bytes	16 MiB

A specific member cap never exceeds max_member_bytes.

from mlbundle import ArchiveLimits, inspect

strict = ArchiveLimits(
    max_archive_bytes=16 * 1024 * 1024, max_structure_bytes=4 * 1024 * 1024
)
view = inspect("trained.mlbundle", limits=strict)

Values above a ceiling are rejected by validation. Pass the same object to export, inspect or load; an upload service would typically lower max_archive_bytes to its request size.

Fixed ceilings

Item	Ceiling
JSON nesting depth	32
Features	100,000
Classes per output	100,000
Outputs	1,024; XGBoost, CatBoost and ONNX: 1
user_metadata	1,024 top-level keys, 64 KiB, depth 8
Hyperparameters	4,096 keys, 1 MiB, depth 16
XGBoost	100,000 trees, depth 14
Record string values	64 characters

What limits do not cover

These ceilings bound archive parsing and the native state mlbundle validates. They do not cap process memory, prediction rows or CPU time during inference. A runtime must enforce those itself.

The normative container limits are in Container and limits, native work ceilings in Prediction state and ONNX bounds in ONNX mapping.