Document conversion tool docling adds direct extraction of images embedded in PDFs and a speech recognition engine for Apple silicon. Developers automating conversion gain new processing options. Invalid model settings previously ignored on the command line now produce errors or warnings.

1. Specifying a model selects the processing path

Released on November 8, 2026, this version automatically selects the corresponding processing path when a vision-language model is specified. These models read page images and turn them into text. Previously, model settings were ignored unless a processing path was selected separately. Now, explicitly pairing a model with an incompatible path causes an error. With no options, the standard path remains in use.

For folders containing mixed formats, users can specify both model types without selecting a processing path separately. PDF files and images then use the vision-language model, while audio uses the speech model. Assigning a speech model to PDFs produces a warning while continuing the existing conversion. Supplying only PDFs to a speech-only path produces an unsupported-format error.

The new rules expose invalid combinations earlier but may change implicit behavior that existing automation relies on. Command-line automation needs a review of pipeline and input format combinations.

2. PDF picture items can use embedded images

Standard PDF conversion usually renders a page as an image and crops picture regions. With the new optional feature enabled, it directly extracts an embedded image when a region contains a single image that can be safely extracted. This preserves the original pixel dimensions, avoiding quality loss from page rendering and cropping. The pypdfium2 and docling_parse backends, the components that perform the conversion, support this feature.

Matching checks whether the image and picture region overlap by at least 80%. The author described this threshold as arbitrary but effective in testing. Regions that are hard to extract directly, such as those mixing text and vector graphics drawn with lines and shapes, use the existing cropping method. The change applies to picture items.

3. mlx-audio handles speech recognition on Apple silicon

The new mlx-audio backend runs speech recognition on Apple silicon without Torch. Some supported models select this engine automatically based on hardware, installed packages, converted model availability, and device settings. These models use Transformers when automatic switching conditions are not met. Presets that pin them to the existing engine have also been added.

Granite speech models do not return timing information directly. Both engines now split recordings at pauses into roughly 30-second chunks, ranging from 20 to 40 seconds, and attach timestamps to each chunk. A Transformers issue that truncated text by applying the generation limit to the entire recording has also been fixed.

According to the change proposal, Granite Speech 5.0 2B reported success under the previous 256-token generation limit despite retaining only 1,227 of about 3,300 characters from a 3-minute recording. The new default limit is 1024 and applies to each chunk. This change aims to reduce missing output when processing long recordings.

The broader choice comes with a larger installation footprint for docling[asr], which requires macOS 14 or later on Apple silicon. If only format support is needed, docling[audio] can also be installed on macOS 13.

The docling models command shows the engines available for each model and the engine automatically selected on the current device. It can help verify the execution path before adopting the new engine.

4. An OCR-specific model joins the vision-language path

A new preset connects MiniCPM-OCR 1.0, a 1.3-billion-parameter model for recognizing text in English and Chinese documents. It returns full-page text and Markdown rather than word or line coordinates, so it is integrated into the vision-language model path.

It also includes repetition-prevention settings from the model card. The change proposal says that without these settings, text repeated on dense test pages, while the same two-page PDF completed correctly after they were applied. Local execution requires Transformers 5.21.0 or later.

In Python, create pipeline options from the preset as follows, then connect them to the converter.

from docling.datamodel.pipeline_options import VlmConvertOptions, VlmPipelineOptions

pipeline_options = VlmPipelineOptions(
    vlm_options=VlmConvertOptions.from_preset("minicpm_ocr"),
)

5. Document parser conversion errors are also fixed

Markdown now interprets only valid character references ending in semicolons, the notation that converts symbols into characters. Strings containing backslash-escaped special characters also remain in one item with the surrounding text.

In JATS, a scholarly document format, tables included as images and their captions are preserved. Footnotes in regular tables are also linked.

HTML attaches captions to their corresponding images and merges code tags split within a single block. Preserving document structure reduces cases where captions or code become detached.

This release's practical value lies in making model selection determine the actual processing path and preserving original images and text from long recordings more fully.