This model is not a single-shot end-to-end parser. It runs as a pipeline:
detect layout, crop each region, then call the model again per region with a
task-specific prompt.
Prompts
Each task uses a fixed prompt (note the leading newline):
Task
Prompt
Layout
\nLayout Detection:
Text
\nText Recognition:
Table
\nTable Recognition:
Formula
\nFormula Recognition:
Figure
\nImage Analysis:
Table recognition returns OTSL. Other output formats are left to the caller.
Inference notes
Serve with --trust-remote-code and --limit-mm-per-prompt '{"image":1}' (one image per request).
Set enable_thinking=False in the chat template.
Pass skip_special_tokens=False when decoding.
Greedy decoding (temperature=0).
Feed page images, not PDFs. Chat UIs (e.g. open-webui) with free-form prompts
will not work — use the prompts above.
Benchmark methodology
The ParseBench score is an end-to-end pipeline measurement — this model served via vLLM plus deterministic rule-based post-processing of model output — consistent with how all ParseBench providers are evaluated (every provider is a submitter-hosted endpoint). No other learned models, classifiers, or ensembles are involved: single model, single pass.
About
Built by KoreaDeep, a document-AI company. The larger KDL-Frontier-Parser-ultra is available through DEEP Agent.
