Datalab Marker v2 vs MinerU, Docling, and Liteparse: Benchmark Breakdown

Datalab just shipped Marker 2, and it's not a patch job — it's a full rebuild around three pieces the team had been quietly shipping for months: Surya OCR 2, a small 20-million-parameter layout model, and a rebuilt pdftext that runs three times faster than the old one. The headline number comes from olmOCR-bench, a third-party benchmark built by Allen AI out of 1,403 PDFs and roughly 8,400 pass/fail checks covering things like math rendering, table structure, and reading order. Marker's balanced mode hits 76.0% overall and 83.5% on born-digital PDFs, while sustaining 2.9 pages per second on a single B200 GPU.

That throughput figure only makes sense once you understand what changed underneath. Marker 2 exposes three separate conversion paths instead of forcing one workflow on everyone. Balanced mode uses the Surya VLM for layout and re-OCRs pages when the embedded text looks bad — it's the slowest but most accurate option, and the one that actually needs a GPU. Fast mode swaps in a lightweight rf-detr/onnx detector with pdftext and touches the VLM only when necessary, scoring 66.6% at a fraction of the cost. And there's a disable_ocr mode that skips the VLM entirely, runs on pure CPU, and still manages 43.6% at 23.7 pages per second.

The architecture behind the speed is arguably the more interesting change. Marker now runs many thin CPU workers against a single shared Surya inference server, with the parent process budgeting VLM concurrency across all of them. Throughput scales with server capacity rather than how much VRAM one process happens to have, which is why Datalab reports balanced mode sustaining roughly 2.9 pages per second versus about 0.3 pages per second single-stream on the same hardware.

Against the competition, the gaps aren't subtle. MinerU's pipeline backend — the closest architectural match — scores 72.7% overall, essentially tied with Marker on born-digital documents (83.3 vs 83.5), but Marker's balanced mode runs 5.4 times faster. Docling, run on its default text-layer-plus-OCR pipeline, trails further at 50.3% while also being slower at 2.1 pages per second. LiteParse, the Rust-based parser from the LlamaIndex team, plays an entirely different game: with OCR off it claims 1,721 pages per second but only 20.4% accuracy, versus Marker's CPU-only 43.6%. None of this touches the full-page VLM tier — Datalab's own Chandra 2 scores 85.8% and Gemini Flash 3.5 hits 76.4%, tools Datalab frames as a separate category rather than direct competitors to a pipeline like Marker.

Licensing is where the four tools actually split apart for anyone shipping a product. Marker's code is Apache 2.0, but the model weights sit under a modified OpenRAIL-M license that's free for research and for startups under $5 million in funding or revenue — past that, a commercial license is required. MinerU's own open-source license stays free up to 100 million monthly active users or $20 million in monthly revenue. Docling is straightforward MIT. LiteParse is open source too, with LlamaParse sitting alongside it as the paid option for harder documents.