Deploy MinerU 2.5 Pro on RunPod in two clicks
Open-source RunPod Serverless template. Your endpoint scales to zero; best-case warm parses measured about $0.0003 per page on a 24 GB RTX 4090.
How it works
This is an open-source repo and RunPod Hub template — not a hosted service. Deploy it from the Hub (or fork it for full control), and the worker runs on your RunPod account against your wallet.
Under the hood it is one Docker image wrapping MinerU 3.4.x with the MinerU2.5-Pro-2605-1.2B VLM by default. Send a PDF as a URL, base64 blob, or path on a mounted volume. Get back Markdown, a structured content_list, the raw middle.json, and extracted images. When traffic stops, RunPod tears the worker down in seconds.
What you get
MinerU 2.5 Pro output
The 1.2B VLM produces Markdown, tables, formulas, reading order, and structured JSON. See the model card for benchmark scope and limitations.
Pay per second, not per hour
Best-case warm parsing measured about $0.0003 per page on a 24 GB RTX 4090. Content density and cold starts raise the real cost; see the measured range.
No glue code
Deploy from the Hub in one click, or fork and let RunPod auto-build your copy. Either way: ten minutes from sign-up to first parse.
Published license terms
MinerU uses an open-source license based on Apache 2.0 with additional terms and stated commercial thresholds. Review those terms for your use case.
Try it in 30 seconds
from mineru_client import MineruClient
client = MineruClient(endpoint_id="<your-endpoint-id>")
result = client.parse_document(file_url="https://example.com/report.pdf", end_page=4)
client.save_tarball(result, "./out/doc")
# → markdown + content_list + middle.json + imagesPick your starting point
How it's built
Architecture, what's in the repo, supported workloads, and how it compares to Marker, GROBID, and Nougat.
Deploy in 10 minutes
Sign up for RunPod, deploy mineru-runpod from the Hub, paste the endpoint id, parse your first PDF.
Pick the right GPU
When 24 GB is enough, when to jump to 48 GB, and which RunPod pool IDs map to which workload shapes. Includes the official MinerU hardware compatibility table.
When something breaks
Hub build flakes, Blackwell crashes, Latin-on-Cyrillic, OOM, cold starts. How to read the debug block to diagnose anything else.
Observe production jobs
Export structured logs, traces, and GPU metrics to any OTLP/HTTP backend with the OpenTelemetry exporter already included in the image.
The source
The wrapper code is MIT-licensed; MinerU and its model retain their own terms. Issues and PRs are welcome.
Blog
Project notes and announcements.