NuMarkdown 8B Thinking

NuMarkdown-8B-Thinking is an image-to-text model that converts documents into clean Markdown files, optimized for retrieval-augmented generation (RAG) applications. It is a reasoning vision-language model (VLM) that generates "thinking tokens" to analyze document layout before producing the final Markdown output, with thinking token volume ranging from 20% to 500% of the answer depending on task difficulty.

The model is a fine-tune of Qwen 2.5-VL-7B, trained in two phases: supervised fine-tuning on synthetic reasoning traces from public PDFs, followed by reinforcement learning (GRPO) with a layout-centric reward on challenging image examples. It is released under the MIT license and is compatible with Text Generation Inference, Inference Endpoints, and Azure deployment (US region). As of its release, the model has received over 1,800,000 downloads and 477 likes on Hugging Face, indicating strong community adoption.

Benchmark Results

In an arena ranking with approximately 500 model-anonymized votes using a trueskill-2 system, NuMarkdown-8B-Thinking (labeled NuMarkdown-reasoning) achieved a μ score of 26.10 (σ 0.79), outperforming OCRFlux-3B (24.63), GPT-4o (24.48), and non-reasoning variants, while remaining competitive with Gemini Flash reasoning (26.75). A win/draw/loss matrix against other models based on image-only evaluation is shown below.

Example Output

The model demonstrates capability on documents with complex layouts, including multi-section headers, bullet points, tables with merged cells, and footnotes, as illustrated in the generated output below.