MinerU 4.x Parsing Tiers: A Practical Guide

[myhloli]

Tip

Start with standard for most PDF and image parsing tasks.
Choose basic when CPU compatibility and resource usage are your main concerns, and flash when you only need to extract content quickly. If the results from standard clearly fail to meet your requirements, try advanced on the documents that need further attention.

1. From multiple parsing modes to one model with four tiers

MinerU 4.x replaces the implementation-oriented names used in 3.x, such as pipeline, hybrid (medium / high), and vlm, with a unified model identifier: mineru-<version>. Four parsing tiers let you choose the balance between quality and cost:

flash → basic → standard → advanced

Here, “one model, multiple tiers” refers to a unified product interface and model identifier. Internally, each tier combines native parsing rules, conventional small models, and vision-language models (VLMs) as needed. When choosing a tier, focus on your quality requirements, available resources, and acceptable processing time.

As a rough analogy, you can think of these tiers as reasoning-effort levels in an LLM:

MinerU tier	Reasoning-effort analogy	Main purpose
flash	low	Extract content quickly with as little inference as possible
basic	medium	Use conventional small models to balance baseline quality and device compatibility
standard	high	Combine small models and a VLM for the best balance of parsing quality and cost
advanced	max	Spend more on VLM inference to tackle difficult cases and specific requirements

This analogy describes effort and trade-offs. Lower tiers are generally faster and use fewer resources. Higher tiers can handle more complex documents, but parsing accuracy does not necessarily improve with every increase in tier, especially when moving from standard to advanced.

The four-tier comparison in this guide mainly applies to PDFs and images. Supported native document formats, such as Office files, use flash. For these formats, a lower tier does not imply lower accuracy. See Tiers and Runtimes for supported formats and interface-specific behavior.

2. Understanding speed, accuracy, and cost

The illustration below shows the trade-offs between the four tiers. It is not based on measured benchmark results and does not compare MinerU with other document parsing tools.

3. Why we recommend standard as the starting point

standard is our recommended tier for everyday parsing and the sweet spot between quality and cost. It aims to deliver accuracy close to full-page VLM parsing in most scenarios, while keeping resource use and processing time closer to basic.

As with LLM reasoning levels, the highest setting is not necessarily the recommended default. OpenAI's reasoning model documentation likewise recommends evaluating higher reasoning effort on your specific tasks to determine whether the improvement justifies the extra latency and cost.

For MinerU, you should not expect advanced to substantially outperform standard in most common scenarios, or send every task to it simply because it is the highest tier. Its additional value is mainly in edge cases: when a small model makes a significant layout error, a VLM's greater ability to generalize may improve the result or even solve the problem outright.

If your goal is to improve parsing accuracy, we recommend trying advanced only after reviewing the results and finding standard clearly unacceptable for your use case, and only if you are willing to spend substantially more time and compute. Compare the results on the problematic pages first. If the improvement is meaningful, you can then decide whether to use it more broadly.

If you specifically need image descriptions or chart interpretation, you can also choose advanced for those capabilities, as described below.

4. What each tier does

flash: Primarily rules and native content extraction

flash aims to rely on deterministic rules and native document structure wherever possible, minimizing model inference. For scanned PDFs, images, and other inputs without usable native text, it uses OCR as needed to extract content.

It is primarily intended for two types of tasks:

• Parsing supported native document formats, such as Office files.
• Quickly extracting content from well-structured digital PDFs with a usable text layer, for previews, search, or further processing.

For PDFs, flash is intended as a fast content extraction tool, with no guarantee of layout or content accuracy. If reading order, complex formulas, table structure, or content completeness is critical, use basic or standard and review the results.

basic: No VLM dependency, with an emphasis on CPU compatibility

basic uses the necessary conventional computer vision (CV) small models and does not depend on a VLM. Its main components include:

• Layout: Layout detection and region segmentation.
• MFR: Mathematical formula recognition.
• Table: Table recognition and structure reconstruction.
• OCR: Text detection and recognition.

Lightweight inference options such as ONNX support a broad range of CPU devices, making this tier suitable for limited-resource environments and simpler deployments. A Torch backend for small models is also supported; the inference backend and the parsing tier are separate choices.

For PDFs with a usable text layer, basic extracts embedded native text wherever possible to avoid unnecessary OCR.

standard: Apply VLM inference where it adds the most value

standard aims for the best balance of parsing quality and cost. Small models handle layout detection, while a VLM handles formulas, tables, and text regions that require recognition, improving results for complex content that conventional small models struggle with.

For PDFs with a usable text layer, it also makes use of embedded native text wherever possible, reducing redundant recognition and unnecessary model generation.

The goal is accuracy close to conventional full-page VLM parsing at a cost closer to basic. This is a design goal; actual processing time and quality gains depend on the document and runtime environment.

If you have no specific resource constraints or special requirements, start with standard.

advanced: Use VLM inference more broadly and spend more compute on difficult cases

advanced assigns more parsing stages to a VLM, spending more compute and time on difficult scenarios. It also extracts embedded native PDF text when available to help reduce hallucinations from model-generated content.

The main difference from standard is that advanced also uses the VLM for layout analysis. This significantly increases inference cost. However, a VLM's ability to generalize may produce better results on unusual layouts or documents where small models make incorrect predictions.

advanced also enables the following capabilities by default:

• Image descriptions (image): Describe image content in text.
• Chart parsing (chart): Attempt to extract or describe information in bar charts, line charts, and similar visualizations.

These capabilities can help turn visual information into searchable text, but the output may be inaccurate. In particular, verify numerical values and relationships extracted from charts. Here, chart refers to data visualizations; it is distinct from the table parsing already supported by basic and standard.

advanced and standard share model resources. Switching tiers does not require downloading a separate “Advanced model.” The additional cost comes primarily from differences in the parsing workflow and the amount of VLM inference performed.

5. Practical recommendations

Your requirement	Recommended starting point
Parse supported native documents, such as Office files	flash
Quickly extract content from digital PDFs, accepting possible layout and content errors	flash
Run mainly on CPUs and balance baseline quality with lower resource usage	basic
Get high-quality results for most PDFs and images	standard
standard clearly falls short on specific documents, and you are willing to spend more to try to improve them	advanced; compare the problematic pages first
Generate image descriptions or interpret chart content	advanced; manually verify the output

Use --tier to select a tier with the local CLI:

# Everyday parsing
mineru-kit parse input.pdf -o output-standard.md --tier standard

# Try advanced if standard is inadequate, keeping both outputs for comparison
mineru-kit parse input.pdf -o output-advanced.md --tier advanced

When investigating a specific page, add --pages 3 to both commands to compare page 3 first and avoid processing unrelated pages.

In the API or WebUI, the available tiers depend on what the service has enabled. auto is an automatic selection strategy, not a fifth parsing tier. Whether the server enables advanced and which tier an individual request selects are separate decisions.

When reviewing results, check for missing text, incorrect reading order, inaccurate formulas, and broken table row or column relationships. If advanced still does not meet your requirements, please share the original sample or a minimal reproduction file, together with your MinerU version, selected tier, affected page numbers, actual output, and expected result. This helps us identify the problem.