Docling vs safetensors

Docling converts an unusually wide set of document formats into a single structured representation, and the release train is dense: nine releases in a month, most carrying one or two new capabilities under a long tail of backend fixes. The recent work splits cleanly in two directions. Format reach keeps extending outward (Outlook .msg, EBCDIC, legacy binary Office formats, video), while the internals are being pulled apart into selectable components: v2.120.0 exposes --layout-engine and --table-structure-engine on the CLI, and the OCR layer was refactored to resolve PP-OCR languages by version and backbone. Parsing fidelity work is concentrated in docx, pptx and odf, where reading order and list structure are still being corrected release over release.

The engine layer is where the interesting movement is. Docling is shifting from one opinionated pipeline to a set of interchangeable layout, table and OCR backends the caller picks per run, which turns the library into a harness for models rather than a fixed parser. A second thread worth watching: the project shipped agent skills for itself in v2.118.0 and added uvx installation docs for them in v2.120.0, alongside a separate docling-client package, all of which point at being consumed programmatically by agents rather than only imported as a Python library. The heading-level inference from font weight, slant and case in v2.120.0 shows the other half of the strategy, extracting structure from typography rather than from markup.

Expect the --layout-engine and --table-structure-engine selection to spread from the CLI into the service API, which already gained heading-level inference and chunking options in the last two releases. The agent-skills and docling-client threads are too new across two releases to call a direction with confidence.