PatSnap trains a 0.3B OCR model for patents and matches MinerU 2.5 on benchmarks
Brilliant_Rich3746 · reddit · 2026-07-28
PatSnap says it trained MOSS-OCR, a 0.3B OCR model optimized for patent documents.
- The model was trained from scratch on 50M+ samples, with patent-domain data as the core.
- It is block-level only: layout detection and reading order are handled by a separate pipeline, which keeps the model small and fast.
- On OmniDocBench v1.6, it matches MinerU 2.5’s overall score of 94.46 at about 25% of the parameters.
- On the team’s own patent-bench, it ranks first at 93.49.
- They report about 59 QPS on a single RTX 4090 via vLLM, roughly 2× MinerU 2.5’s throughput.
- The repo is Apache 2.0, and a companion service called Hiro-Smart-Doc wraps layout detection plus OCR into one FastAPI pipeline.
More from Apps
- PostHog turns its hedgehog mascot into a “product self-driving” meme — thedealdirector · 2026-07-28
- Codex automation checked SF apartment listings hourly and won the lease first — nickbaumann_ · 2026-07-28
- Google AI Overviews rise from 15% to 43% of searches in a year — HaktanSuren · 2026-07-28
- Kimi K3 Hits LM Studio: 2.8T Parameters with 1M Context — gnukeith · 2026-07-28
- Neo4j Aura Turns PDFs into Queryable Knowledge Graphs — JeremyCMorgan · 2026-07-28
- A prototype rethinks coding with agents as a coordinated, parallel workflow — Wattenberger · 2026-07-28