Datalab's Marker v2 parses documents at 23.7 pages/s on a single B200 GPU, per @akshay_pachaar. The open-source pipeline handles PDFs, images, DOCX, and PPTX into clean markdown.
Key facts
- 23.7 pages/s on a single B200 GPU
- 90+ languages supported
- 100% open-source license
- Handles PDF, image, DOCX, PPTX
- Output: clean markdown
Parsing one PDF is trivial. Parsing millions is a throughput problem, and Datalab's Marker v2 is built for that scale. The new pipeline, announced via @akshay_pachaar's post, claims up to 23.7 pages/s on a single NVIDIA B200 GPU — a pace that would chew through a 10,000-page corpus in about seven minutes per GPU.
Marker v2 isn't just a speed bump; it's a language expansion. The tool supports 90+ languages, which matters for enterprises processing multilingual contracts, invoices, or academic papers. The output is clean markdown, a format that feeds directly into RAG pipelines, LLM fine-tuning sets, and knowledge-base ingestion workflows.
The 100% open-source license is the notable structural move. Datalab is betting that developer adoption and community contributions will outflank closed rivals like Adobe's Extract or AWS Textract, which charge per page and lock output into proprietary schemas. Open source also means teams can self-host on their own B200 clusters, avoiding per-call API fees at scale.
Why throughput matters more than accuracy here
Most parsing tools benchmark on accuracy — table structure, formula extraction, layout fidelity. Datalab's headline metric is raw speed, which signals a different target customer: the bulk-ingestion team that needs to convert a legacy document lake into structured text before an LLM can touch it. At 23.7 pages/s, the bottleneck shifts from parsing to storage and downstream embedding.
The B200 comparison is also deliberate. Marker v2 is optimized for the current flagship data-center GPU, not a consumer card. That ties the tool's performance to the AI infrastructure buildout — teams already renting B200s for training can repurpose idle capacity for document parsing overnight.
The source post does not disclose benchmark methodology, hardware configuration details, or accuracy numbers. Datalab has not published a technical report or comparative evaluation against prior versions or competitors. The 23.7 pages/s figure is a vendor claim, not an independent measurement.
What this means for RAG pipelines
For teams assembling retrieval-augmented generation systems, Marker v2 lowers the barrier to high-quality source documents. Markdown preserves headings, lists, and code blocks better than raw text extraction, which improves chunking and embedding quality. The 90+ language support also opens up non-English corpora that many parsing tools handle poorly.
The open-source license means no per-page pricing. At scale, that's the difference between a $0.0015/page API bill and a one-time infrastructure cost. For a 10-million-page archive, that's $15,000 saved per pass — and the ability to re-parse whenever the model improves.
Datalab hasn't announced a changelog or migration guide for Marker v1 users, but the v2 tag implies a breaking or major update. Teams on v1 should test their document mix before upgrading.
What to watch
Watch for independent benchmarks on standard parsing datasets like PubLayNet or DocBank, and whether Datalab publishes accuracy numbers to match the speed claims. Also track adoption in open-source RAG stacks — if LangChain or LlamaIndex add Marker v2 as a default loader, that's a stronger signal than any tweet.
Key Takeaways
- Datalab's Marker v2 parses documents at 23.7 pages/s on a B200, supports 90+ languages, and is open-source.
- The speed claim targets bulk ingestion for RAG pipelines.
