Marker v2: 23.7 pages/sec on one B200 GPU

Marker v2 tackles a classic bottleneck in document parsing: GPU underutilization. Traditional scaling adds workers, each with a private model copy, leading to idle time and redundant memory. By centralizing inference on a shared server, Marker v2 ensures the GPU processes many pages in one pass, achieving 23.7 pages per second on a single B200 with OCR off. This design also decouples throughput from VRAM, making scaling more cost-effective for businesses.

The benchmark results provide concrete evidence of its efficiency. On olmOCR-bench, balanced mode scores 76.0% against MinerU's 72.7%, while running over 5x more pages per second. However, the mode choice is critical: fast mode sacrifices math accuracy because it reads equations from the text layer instead of the page, so math-heavy corpora should stick to balanced. This trade-off is typical of optimization strategies that prioritize speed over fidelity.

For non-technical readers, the takeaway is that Marker v2 offers a practical way to reduce parsing costs and improve throughput without sacrificing quality, provided the right mode is selected. The open-source nature allows adoption without vendor lock-in, though teams should validate performance on their own document types.