Hybrid Backend Effort Levels#

MinerU's hybrid backend exposes two parsing effort levels — medium and high — that trade inference speed for extraction accuracy. The valid values are defined as HYBRID_ANALYZE_EFFORTS = {"medium", "high"} 1, with medium as the default 2.

What Each Level Does#

The effort level controls which VLM extraction path is invoked per page window 3:

Medium effort feeds pipeline layout detections into the VLM via _build_medium_vlm_layout_blocks, bypassing the VLM's independent layout analysis. Image and chart analysis is unconditionally disabled at medium effort by _resolve_effective_image_analysis, which substantially reduces peak memory for image-heavy documents 4.

High effort calls batch_two_step_extract (or the async variant aio_batch_two_step_extract), letting the VLM run its own full two-step layout + extraction pass without layout pre-seeding.

Both paths share the same post-processing: OCR det sidecars, inline formula merging, and title splitting via _apply_layout_title_split 5.

Accuracy Comparison#

6

Medium effort suits speed-sensitive pipelines or memory-constrained deployments. High effort is recommended when document fidelity is the priority, particularly for engineering drawings, densely-ruled tables, or non-standard layouts 7.

CLI and API Usage#

The effort level is set via the --effort / -e flag on the hybrid backend. Validation is handled by validate_effort at the CLI layer and _validate_parse_effort inside hybrid_analyze.py.

mineru -p input.pdf -b hybrid-engine --effort high

The online SaaS API is locked to medium effort and does not expose the --effort parameter; self-hosted deployments may use either level freely 7.

Historical Note: "Flash Mode" / "Hybrid-Flash"#

Prior to the current design, an early proposal (PR #5090, closed without merge) introduced a separate hybrid-flash-engine backend alongside new constants BACKEND_HYBRID_FLASH_ENGINE and BACKEND_HYBRID_FLASH_HTTP_CLIENT 8. This "flash mode" was a discrete backend variant rather than a parameter on the existing hybrid backend.

That approach was superseded. The subsequent PR #5236 (merged 2026-07-06) consolidated terminology, replacing the flash backend concept with the --effort parameter on hybrid-engine 9. flash references no longer appear in public-facing options; legacy backend aliases are handled by normalize_backend.

Key Files#