What happened
Nanonets describes OCR2-3B as part of its OCR2 family, alongside OCR2-Plus and OCR2-1.5B-exp. The Hugging Face model page identifies Qwen2.5-VL-3B-Instruct as its base model. Its headline workflow takes a document image and asks the model to extract text in reading order while preserving useful structure. The suggested prompt asks for tables in HTML, equations in LaTeX, image descriptions in <img> tags, watermarks and page numbers in tags, and checkboxes as Unicode symbols.
The feature list extends beyond ordinary OCR. The page says the model can distinguish inline and display math; describe charts, graphs and logos; isolate signatures; extract watermarks; represent complex tables in Markdown and HTML; and express flow charts and organisational charts as Mermaid. It also says the model was trained on handwritten documents across multiple languages, naming English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean and Arabic among others. Its visual question-answering behavior is described as answering directly when the information appears in the document and returning “Not mentioned” otherwise.
For local use, the Transformers example loads the model with AutoModelForImageTextToText, an auto processor and tokenizer, and sets device_map="auto" and attn_implementation="flash_attention_2". It shows image input through a chat template and deterministic generation with do_sample=False. A separate example serves the model with vLLM and calls it through an OpenAI-compatible client; that sample sets temperature to 0.0 and allows up to 15,000 output tokens. The page also points to Docstrange’s extraction API and hosted demo. These examples describe possible routes, not a statement that a hosted inference provider currently deploys the model: the page says it is not deployed by an Inference Provider.
The evaluation tables report different comparisons for OCR2-Plus and OCR2-3B. On the listed VQA results, OCR2-3B scores 78.56 on ChartQA and 89.43 on DocVQA; the same table gives Gemini 2.5 Flash 84.82 and 85.51, respectively. The page’s Markdown comparisons give win, lose and both-correct rates, but those rates are relative to the named Nanonets model in each table. Nanonets also recommends increasing image resolution to improve performance, trying repetition_penalty=1 for complex tables, and using Docstrange’s “Markdown (Financial Docs)” option for table-heavy financial documents.
What the archive adds
New Runtime’s earlier coverage includes Xiaomi Research’s 0.8B document-parsing and OCR model from November 5, 2026. OCR2-3B is a new event, not a restatement: it adds a separate model with a different stated base model, workflow examples and feature set to the document-OCR thread.
Trend signal
This does not yet fit a pattern New Runtime tracks. OCR2-3B is a document-processing release, but the available facts do not establish a broader shift in agent architecture, runtime security, interoperability, or the economics of completed tasks.
Why now
The release page is directly usable for teams evaluating document ingestion: it includes loading examples, inference options, prompts and task-specific suggestions rather than only announcing a model. That makes the immediate timing practical for a team already choosing or revisiting its OCR stage, though the page does not explain why Nanonets released it at this particular moment.
Why it matters
Document pipelines often need more than a flat string. Tables, page boundaries, equations, checkboxes and chart descriptions can carry the meaning that a downstream model or workflow needs. A model that emits explicit structure gives teams a candidate route from scanned pages to data their systems can process, inspect and pass onward.
The useful question is whether its output is reliable on the documents a team actually handles. Nanonets’ examples make a first trial straightforward, and its recommendations point to variables worth testing: image resolution, table complexity and output format. The reported VQA scores also show that performance varies by task and comparator; they should not be reduced to a universal ranking. Teams should evaluate extraction errors against their own tolerance for missed fields, malformed tables or incorrect formulas before replacing an established pipeline.
Limits and uncertainty
The page’s metrics and feature claims are published by Nanonets; they are not independent validation. The evaluation tables do not, in the excerpted presentation, fully establish dataset composition, sample sizes, scoring protocol or confidence intervals, so small differences should be treated cautiously. The page gives no latency, throughput, hardware requirement, deployment cost or production error rate. Nor does a feature list quantify performance on difficult scans, handwriting, mixed layouts or sensitive documents. The model is identified as based on Qwen2.5-VL-3B-Instruct, but the page does not establish a license in the text presented here. Teams should verify current model-card terms and test outputs before relying on it operationally.
Skill to build now
Build a small evaluation set from representative pages in your own pipeline: include a dense table, a form with checkboxes, a page with equations, a scan with handwriting and a chart. Run the model with the page’s suggested extraction prompt, then compare the structured output with a human-checked reference. For table-heavy financial pages, test the suggested financial-docs output mode separately; record field errors and malformed structure rather than judging only whether the prose looks plausible.
Reader takeaway
OCR2-3B is worth treating as a candidate structured-document parser, not as proof that document understanding is solved. Its examples lower the cost of a trial; only evaluation on your own pages can show whether its Markdown, tables and semantic tags are dependable enough for downstream work.