LlamaParse is a hosted, proprietary parsing API from LlamaIndex that costs $1.50 to $37.50 per 1,000 pages across its four tiers at pay-as-you-go rates, with self-hosting only on Enterprise plans. Docling is an MIT-licensed open-source library you run yourself for the cost of compute, and its default layout and table models are also under permissive licenses. Marker is Apache 2.0 code from Datalab whose model weights are free for research, personal use, and companies under $5M in revenue and funding. If you want a self-hosted LlamaParse alternative with no license threshold, that is Docling; if you want no infrastructure at all, that is LlamaParse or a hosted Docling API.
Quick answer: LlamaParse vs Docling vs Marker
| LlamaParse | Docling | Marker | Parsebridge (ours) | |
|---|---|---|---|---|
| License | Proprietary SaaS | MIT code; default models Apache 2.0 / CDLA-Permissive-2.0 | Apache 2.0 code; model weights under a modified OpenRAIL-M license | Proprietary service; the parsing is done by MIT-licensed Docling |
| Self-hosting | Enterprise plans only (Kubernetes, BYOC) | Yes, Python library or docling-serve | Yes, Python library; balanced and vllm modes add a local inference server | No; self-host Docling instead |
| Hosted option | LlamaParse cloud (North America, EU) | Third-party APIs such as Parsebridge | Datalab API (runs Datalab’s Chandra and other models) | Yes, it is a hosted Docling API (file upload or URL) |
| Pricing | $1.50 per 1,000 credits pay as you go ($1.25 on the $500/month Pro plan), 1 to 25 credits per page by tier; 10,000 free credits/month | Free; you pay for compute | Free under $5M revenue and funding; above that, a Datalab license | From $17 a month for 5,000 pages, $3.40 to $0.86 per 1,000 pages across plans; 50 free pages once |
| OCR and scanned PDFs | Built in, with LlamaIndex’s OCRia models from the Cost-effective tier up; docs recommend the Agentic tier for scanned pages | Pluggable: EasyOCR, Tesseract, RapidOCR, macOS Vision, Nemotron OCR, or a VLM | In-process ONNX OCR in fast mode; Surya VLM via vLLM (NVIDIA) or llama.cpp in balanced and vllm modes | Docling OCR, turned on automatically when a PDF has scanned pages |
| Tables | Markdown or HTML tables, optional XLSX export | TableFormer model | Rebuilt from the text layer, hard tables sent to the VLM in balanced mode, optional LLM mode | Docling TableFormer in accurate mode |
| Output | Markdown, text, JSON items, XLSX, rendered PDF | Markdown, HTML, lossless JSON, DocTags, DocLang, WebVTT | Markdown, JSON, HTML, chunks | Markdown in a JSON response |
| Input formats | 130+ claimed, including audio | PDF, Office, HTML, EPUB, images, audio, video, LaTeX, email, and more | PDF and images; DOCX, PPTX, XLSX, HTML, EPUB with marker-pdf[full] | PDF only, up to 50 MB |
Pick LlamaParse if you want zero infrastructure and can pay per page, especially for a wide mix of file types. Pick Docling if you need to run the parser yourself under a permissive license at any company size, or call it through a hosted Docling API when you don’t want to run it. Pick Marker if you self-host, fit under Datalab’s $5M threshold or will buy a license, and want its Markdown, JSON, and chunk outputs. Pick Parsebridge, our hosted Docling API, if you parse PDFs and want Docling’s output at a fixed monthly price without running it.
There is no overall winner. The right pick depends on licensing, deployment target, and the worst document in your corpus, so test on your own files before you commit.
Hosted Docling API
Parsebridge, which runs this blog, is a hosted Docling API. You send a PDF as a file upload or a URL and get Markdown back in a JSON response, with no servers, GPUs or model downloads to manage. Signing up gives you 50 free pages, and plans start at $17 a month for 5,000 pages. LlamaParse costs $1.50 to $37.50 per 1,000 pages depending on tier at pay-as-you-go rates, and the Parsebridge Basic plan works out to $3.40 per 1,000 pages with MIT-licensed Docling doing the parsing.
Docling
Open-source document parser started by IBM Research Zurich and now hosted by the LF AI & Data Foundation. The code is MIT. The default layout model (docling-layout-heron) is Apache 2.0 and the table model (docling-models) is listed as CDLA-Permissive-2.0 and Apache 2.0; OCR engines and VLMs you add carry their own licenses. Current release is 2.150.0, published November 10, 2026 (PyPI).
from docling.document_converter import DocumentConverter
doc = DocumentConverter().convert("paper.pdf").document
print(doc.export_to_markdown())
Tables run through TableFormer, in accurate mode by default (table options). OCR is pluggable: EasyOCR, Tesseract, RapidOCR, macOS Vision (OcrMac), NVIDIA Nemotron OCR, or a VLM pipeline such as GraniteDocling; the Docling OCR guide has tested code for scanned PDFs. Export formats are Markdown, HTML, lossless JSON, DocTags, DocLang, and WebVTT.
The cost is infrastructure. Docling pulls in PyTorch plus layout, table, and OCR model weights, so it needs a container or a long-lived server. The Docling Docker guide compares the official images and their sizes, the Docling Serve guide covers the REST server, and the serverless guide covers what fits on Lambda and Vercel.
Pick Docling if you need a permissive open-source license, can run a long-lived container, and want one library that covers PDF, DOCX, PPTX, XLSX, and HTML with consistent output. Skip it if you are shipping to serverless and do not want a hosted layer in front.
LlamaParse
Hosted parsing service from LlamaIndex, documented at the LlamaParse docs. The parsing engine is proprietary. The managed service runs in North America and the EU, and self-hosting or BYOC on Kubernetes is available on Enterprise plans.
Since October 2026 the AI tiers run on OCRia, LlamaIndex’s own document parsing model. OCRia-1, announced October 22, is a 4-billion-parameter vision-language model fine-tuned from Qwen3.5-4B. Its technical report says LlamaIndex plans to make a version of it available, and no weights were published as of November 13. The tier docs say OCRia-1 reads the pages on Cost-effective and Agentic jobs unless you set a custom prompt, and the ParseBench-2 post of November 12 says a newer OCRia-1.5 family now powers Cost-effective, Agentic and Agentic Plus. The Fast tier runs no AI model over the page. LlamaIndex says prices did not change with OCRia-1.
The v2 API has four tiers, priced in credits per page. Credits cost $1.50 per 1,000 on Pay as you go and $1.25 per 1,000 on Pro, which gives the dollar columns (list prices as of November 13, 2026, from the pricing docs):
| Tier | Credits/page | $ / 1k pages, Pay as you go | $ / 1k pages, Pro | LlamaIndex’s guidance |
|---|---|---|---|---|
| Fast | 1 | $1.50 | $1.25 | Plain-text documents at high volume; no AI model runs over the page |
| Cost-effective | 3 | $4.50 | $3.75 | Text-heavy documents with simple tables |
| Agentic | 10 | $15.00 | $12.50 | Most real-world PDFs with tables, scanned pages, multi-column layouts |
| Agentic Plus | 25 | $37.50 | $31.25 | Dense financial reports, complex tables, charts, equations |
Layout extraction adds 3 credits per page. The plans are Pay as you go ($0, with 10,000 free credits a month), Pro ($500/month, 400,000 credits a month, plus a limited-time one-time bonus of 800,000 credits for business email domains), and Enterprise. Starter is no longer offered to new customers. On Pay as you go, new jobs stop with a 402 error when the balance reaches zero, until you buy credits or the monthly credits reset.
LlamaParse lists 130+ supported file types across documents, images, spreadsheets, and audio. A job can return per-page Markdown (tables as Markdown or HTML), plain text, structured JSON items, an XLSX workbook of the tables, and a rendered PDF.
LlamaIndex also publishes LiteParse, an Apache 2.0 parser that runs locally. It extracts spatial text with bounding boxes and bundles Tesseract for OCR, without LlamaParse’s models or cloud features. Running LlamaParse itself in your own cloud still needs an Enterprise plan.
Pick LlamaParse if you want zero infrastructure, can pay per page, and parse a wide mix of formats. Skip it if compliance rules out third-party APIs and you are not on an Enterprise plan, or if your volume makes per-page pricing painful at the Agentic tiers.
Marker
Open-source document converter from Datalab. Current release is marker-pdf 2.2.0, published November 12, 2026 (PyPI, GitHub).
pip install marker-pdf
marker_single paper.pdf
Marker 2.2 has four modes. fast, the default on CPU and Apple Silicon, runs layout, reading order, and OCR in process with Datalab’s small Indra ONNX models and needs no inference server. balanced, the default on NVIDIA GPUs, keeps clean text pages on that path and sends tables, math, non-Latin scripts, and scanned pages to the Surya 2.5 VLM. vllm runs the VLM on every page, and no_ocr uses only the PDF text layer. The VLM runs in a local inference server that Marker starts on first use: vLLM in Docker on NVIDIA GPUs, llama.cpp elsewhere. Inline math is converted to LaTeX in balanced mode, and an optional --use_llm mode calls a hosted or local LLM to merge tables across pages and extract form values.
Two licensing details to read before you ship:
- The code is Apache 2.0.
- The model weights use a modified OpenRAIL-M license. They are free for research, personal use, and companies under $5M in both revenue and funding. Above that, commercial use needs a license from Datalab, which does not list a price for it. The license text also rules out use by a company that offers a product or service competing with Datalab’s.
Datalab also sells a hosted API, which runs its Chandra model and other models rather than Marker. As of November 13, 2026, its list price for conversion is $1.33 per 1,000 pages in fast mode, $4 in balanced mode, and $10 in accurate mode, with 25% off if you let Datalab use your data for model training.
Pick Marker if you self-host, are under the $5M threshold or will buy the license, and, for the VLM modes, can run a local inference server next to it. Skip it if you need permissive model licensing at any company size.
Why benchmarks don’t pick the best PDF parser for RAG
Most published comparisons score parsers against DocLayNet (80,863 annotated pages), PubLayNet, DocBank (token-level layout, 500k pages), or FinTabNet (financial tables). These measure layout-box detection or table-cell structure. None of them measure whether a downstream LLM can answer questions correctly from the parsed output.
ParseBench is closer to what RAG cares about. It covers about 2,000 human-verified enterprise pages across tables, charts, content faithfulness, semantic formatting, and visual grounding. It comes from LlamaIndex and reports LlamaParse Agentic as the top-scoring method, so weight the results with that in mind. The same paper scored Docling’s default pipeline, version 2.73.1, at 50.6 overall.
On November 12, 2026, LlamaIndex released ParseBench-2. It roughly doubles the set to about 4,300 cases, adds forms as a sixth dimension next to text, formatting, tables, charts, and visual grounding, and weights each error by how much it would matter downstream, so a wrong number in a table costs more than a change from HTML to Markdown. The dataset and evaluation code are public. In LlamaIndex’s own benchmark, LlamaParse Agentic Plus ranks first of 41 configurations. The published ParseBench-2 leaderboard has no Docling or Marker entries. It does score Datalab’s hosted API, which runs Chandra and other models.
The test that predicts production behavior is your own corpus. Pick five PDFs from your worst cases: a multi-column research paper with equations and references, a financial filing with merged-cell tables and footnotes, a scanned invoice or contract with stamps and rotation. Run each parser, embed the output, ask a dozen questions you already know the answers to, and score whether the answers come back right. For a worked example on five public PDFs, our tested roundup of PDF to Markdown converters runs Docling, Marker, MinerU, PyMuPDF4LLM and MarkItDown side by side and quotes their output.
Docling vs Marker on tables, scans and speed
We ran Docling 2.138 to 2.141 and Marker 2.0.0 on the same PDFs on an Apple M4 Mac mini, with default settings, for the Parsebridge PDF Parser Benchmark 2026. We did not test LlamaParse.
| Test | Docling | Marker 2.0.0, fast mode |
|---|---|---|
| Dense table, 22 rows by 15 numeric columns | 20 of 22 rows right; the last two rows were merged | 22 of 22 rows right; the series names were dropped from the row labels and the year header was missing |
| One-page scanned PDF with a small table | All text and the table, 3.7 to 3.9 s, no extra install | All text and the table, 14.8 s, after installing a llama.cpp binary |
| 60-page born-digital report | 229 to 244 s | 81.3 s |
These numbers come from five PDFs on one machine. Marker 2.2.0 moved fast mode to the in-process Indra engine described above, so its current speed and output can differ from what we measured.
Cost per 1k pages
| Option | $ / 1k pages |
|---|---|
| LlamaParse Fast | $1.50 |
| LlamaParse Cost-effective | $4.50 |
| LlamaParse Agentic | $15.00 |
| LlamaParse Agentic Plus | $37.50 |
| Datalab API, fast | $1.33 |
| Datalab API, balanced | $4.00 |
| Datalab API, accurate | $10.00 |
| Parsebridge Basic ($17 / 5k pages) | $3.40 |
| Parsebridge Growth ($79 / 60k pages) | $1.32 |
| Parsebridge Scale ($259 / 300k pages) | $0.86 |
| Self-hosted Docling or Marker, GPU always-on | Depends on utilization |
Prices are list rates as of November 13, 2026. LlamaParse rows use the pay-as-you-go credit rate; the $500/month Pro plan lowers them by a sixth. Datalab rows are without the 25% data-sharing discount. Parsebridge plans are monthly page allowances, so the per-page figure assumes you use the full allowance.
Self-hosting cost depends on how busy you keep the GPU. An AWS g5.xlarge (one A10G) lists at $1.006/hour on demand in us-east-1 (EC2 pricing), about $734 for a 730-hour month. Spread over 1M pages that is $0.73 per 1,000 pages; over 50k pages it is $14.69, more than every hosted option above except LlamaParse Agentic and Agentic Plus. The 1M case assumes one instance sustains about 0.4 pages/s around the clock, which we have not measured on a g5.xlarge. For reference, the Docling maintainers’ GPU guide reports 3.1 pages/s for the standard pipeline without OCR on a g6e.2xlarge (L40S) on one 192-page PDF.
Pick this when
| Scenario | Choice |
|---|---|
| Data must stay in your infrastructure | Docling or Marker, or LlamaParse self-hosted on an Enterprise plan |
| Permissive license for code and default models at any company size | Docling |
| Company over $5M revenue or funding, self-hosted | Docling, or Marker with a Datalab license |
| Broad format mix, hosted | LlamaParse |
| Serverless deployment that wants Docling output | Hosted Docling API |
| Dense financial reports, hosted, paying per page | LlamaParse Agentic Plus, which LlamaIndex recommends for its hardest documents |
| Embedded in a commercial OSS product | Docling, after legal review of the model licenses |
| High volume with a 24/7 GPU | Self-hosted Docling or Marker |
| Lowest marginal cost at high scale | Self-hosted Docling or Marker, or Parsebridge Scale |
Where Parsebridge fits
Parsebridge runs Docling behind an API, so a Lambda function or a Vercel route can call it without packaging PyTorch or model weights. It takes PDFs up to 50 MB and returns Markdown, uses TableFormer in accurate mode, and turns on OCR automatically when the PDF contains scanned pages. New accounts get 50 free pages, one time, and plans start at $17/month for 5,000 pages. If you already self-host Docling and the deployment works, stay there. If the infrastructure is what stopped you, the hosted Docling API gives you the same parser without running it yourself.