Ant Group Releases PolyOCR, Unified OCR Foundation Models with SOTA Results
GuangJian Team · hf · 2026-10-12
Ant Group's GuangJian team presents PolyOCR-Venus, a family of unified OCR foundation models moving OCR from plain transcription toward general visual intelligence spanning recognition, localization, and reasoning.
Highlights:
- Shared instruction-following framework plus a large-scale data engine converting heterogeneous visual resources into quality-verified OCR supervision
- Competence-Guided Policy Optimization: verifier-based GRPO combined with on-policy distillation via sample-wise routing based on teacher reliability and teacher-student competence gaps
- Releases OCRBench v2.1 with manually verified annotation corrections and task-aligned scoring
Achieves SOTA or highly competitive results on OCRBench v2.1, CC-OCR, OmniDocBench v1.6 and more; open-sourced on GitHub.
More from Models
- Claude Code Users Report Fast Mode Locked Despite Plenty of Remaining Credits — RileyRalmuto · 2026-10-12
- OpenAI's new Decisions API fails a simple coin-flip calibration test — JnBrymn · 2026-10-12
- Unverified leak claims Anthropic's next Claude is being tested in Claude Code, billed as SOTA — Eric520CC · 2026-10-12
- Prediction: Anthropic keeps public model lead until OpenAI's December release — imjustnewatai · 2026-10-12
- Leak claims OpenAI's next models: 5T-param sol on spud, 10T-param astra on doug pre-train — GregCook2011 · 2026-10-12
- Dev vents: OpenAI models still suck at writing, why is that? — beechinour · 2026-10-12