[cau-git]

Add nvidia/Nemotron-3.5-Parse as the VLM preset nemotron_parse_v3_5, with a mapper from its native output to DoclingDocument.

Preset

• Pinned to revision e7ea79242a87ea0ad3919b672bfedb429a683206.
• Engines: Transformers (AutoModel, raw prompt, greedy, repetition_penalty=1.1, max_new_tokens=8744) and the OpenAI-compatible API engine, for NVIDIA's vLLM server (top_k=1, skip_special_tokens=False).
• In-process vLLM is not supported: the model needs NVIDIA's out-of-tree vLLM plugin with exact version pins. MLX is not supported: no MLX port exists.
• Single-pass parsing only (no crop re-parsing).

Remote code consent

• The model needs code from its repository (model_type nemotron_parse is not native in transformers). The preset does not trust it by default.
• New opt-in: trust_remote_code on AutoInlineVlmEngineOptions and the --trust-remote-code CLI flag. Without the opt-in, the Transformers engine raises a clear error that names the repository and revision, before it downloads the model.
• Revision pinning comes from fix(models): warn on unpinned revisions for trust_remote_code model specs #4227. This PR adds the consent gate and extends TestShippedSpecsArePinned to presets that require the opt-in.
• Other presets keep their current behavior; moving them to the gate is a follow-up.

Mapper (docling/utils/nemotron_parse_v3_5_utils.py)

• Content-normalized coordinates map directly to the page (no padding inverse as in 2.0).
• LaTeX tabular tables with \multirow, \multicolumn, \thead and \mbox{} map to cell spans, column headers and rich cells. The LaTeX backend used for 2.0 does not handle these.
• <chart type=...> blocks map to pictures with a chart classification and a TabularChartMetaField (title and data); multi-panel charts give one picture per panel.
• Markdown emphasis and inline <sup>/<sub>/<u> map to formatting runs.
• Captions link to an adjacent table or picture; list items keep their markers; page headers and footers go to the furniture layer; consecutive TOC blocks share one group; text in pictures becomes picture children; Code and Formula use their own item types.

Also included: a fix for a KeyError in the dots inline Markdown parser (setext underline inside emphasis), which the new mapper reuses.

Checklist:

• Documentation has been updated, if necessary.
• Examples have been added, if necessary.
• Tests have been added, if necessary.

Review evidence:

• Before: no preset for Nemotron 3.5 Parse. After: --vlm-model nemotron_parse_v3_5 --trust-remote-code converts pages with the Transformers engine.
• Tests: tests/test_nemotron_parse_v3_5.py, tests/test_hf_model_download_security.py, tests/test_vlm_presets_and_runtime_options.py, tests/test_nemotron_parse_v2.py, tests/test_dots_parser.py: 133 passed, 1 skipped (existing marker). CLI and auto-inline related test files: 370 passed. Pre-commit hooks pass on all changed files.
• Mapper check against NVIDIA's 500 published reference predictions: 10,143 blocks, 375 tables (none empty), 141 charts, 0 errors.

Open points for review:

• Captions with inline formatting (e.g. **Table 1.**) have child runs, and link_adjacent_captions does not link them (same limit as dots).
• The API engine sends the image before the text; NVIDIA's client sends the text first. The chat template only concatenates content, so no effect is expected, but this is not tested.

[github-actions]

✅ DCO Check Passed

Thanks @cau-git, all your commits are properly signed off. 🎉

[mergify]

Merge Protections

🟢 Merge protection satisfied — ready to merge.

Show 1 satisfied protection

🟢 Enforce conventional commit

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

• title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert)(?:\(.+\))?(!)?:

[codecov]

Codecov Report

❌ Patch coverage is 89.25620% with 39 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
docling/utils/nemotron_parse_v3_5_utils.py	88.66%	18 Missing and 16 partials ⚠️
docling/pipeline/vlm_pipeline.py	0.00%	2 Missing and 1 partial ⚠️
docling/cli/main.py	50.00%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

AI triage (advisory)

Checked commit 1d090d3cb2. The analysis read the diff only. It did not run code from this pull request.

The PR adds the nemotron_parse_v3_5 VLM preset for nvidia/Nemotron-3.5-Parse, with a mapper from its native output to DoclingDocument. It also adds an opt-in trust_remote_code option and a --trust-remote-code CLI flag. Small changes to the chandra_utils.py and dots_utils.py helpers support the mapper.

CI safety: no concern found. A maintainer must still approve CI.

2 related pull request(s) in the same code

PR	State	Signal	Assessment
#4789	open	changes the same code	related: This PR adds LightOnOCR-3 presets. It touches the same preset registration and ResponseFormat code, but it adds a different model.
#4908	open	changes the same code	related: Both PRs change _add_inline_text and the Markdown handling in dots_utils.py, with different behavior. They will likely conflict when merged.

_{The ai:* labels are advisory. They do not approve CI and they do not add tests:full. A new push runs this check again.}

[nikos-livathinos]

Some feedback on the way formulas are stored:

• The mapper stores formulas as is (nemotron_parse_v3_5_utils.py:544).
• The model usually writes LaTeX formulas as \(…\).
• The Markdown serializer then adds $$, so we end up with something like:

$$\(f(x)=\sum_{n\in S_x}\frac{1}{2^n},\)$$

This is the same bug as #4903 for Nemotron v2. Please strip the \(…\) of each segment.

Evidence: I run docling locally on a sample from OmniDocBench. Here is a zip with the input image and the
predictions.
nemotron_v3_5_formula_pages.zip

[dolfim-ibm]

lgtm