WeVisDoc-2B
is a compact end-to-end document parsing model developed by Tencent that converts page images directly into structured Markdown output, complete with LaTeX-formatted formulas and HTML tables. Fine-tuned from Qwen3-VL-2B-Instruct, it is purpose-built for document parsing rather than general vision-language tasks, and supports both English and Chinese. Despite its small 2B-parameter size, it delivers near state-of-the-art results, scoring 95.06 Overall on OmniDocBench v1.6—second only to its larger 4B sibling (95.38) and ahead of much bigger models like olmOCR-2-7B, dots.mocr, and Logics-Parsing-v2—while on PureDocBench it achieves a mean Overall score of 73.86 across the three tracks (79.36 Clean, 76.62 Digital Degraded, 65.60 Real Degraded), leading the Clean and Digital Degraded settings among compared parsers. Released under the Apache 2.0 license, it can be served via vLLM (requiring vLLM ≥0.11.1) or run locally with Transformers, and uses the same client tooling as the 4B variant, making it an attractive option for resource-constrained deployments that still require top-tier parsing accuracy.