New layout model benchmarks & recommendations?

[kbumsik]

Question

In the latest release #1910 new layout models are introduced. And the following models are particularly interesting to me:

• ds4sd/docling-layout-heron-101
• ds4sd/docling-layout-egret-xlarge

I have questions regarding those models:

• This seems to be a big update to me, but why they are not documented? Are they still experimental?
• Do you have any benchmark to get to know which model performs better?

Besides questions, thanks you for keeping the project up-to-date!

[dosubot]

Hi @kbumsik! I'm Dosu and I’m helping the docling team.

There isn't any official documentation, benchmarks, or performance comparisons for the new layout models (docling-layout-heron-101 and docling-layout-egret-xlarge) in the repo or changelog as of the latest release. The models are available as configuration options, but they're not marked as experimental or production-ready, and there are no recommendations or published results to help choose between them. The only metrics available are per-page confidence scores (layout_score and ocr_score), which are used internally and not tied to any external benchmarks or comparative evaluations across models. If you want to compare them, you could try both on your own data and monitor those confidence scores, but for now, there's no official guidance or leaderboard to reference.
Source: layout model configs
Source: implementation details

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}