What it is
A single 5B parser turns documents, images, audio, and video into one grounded JSON schema using a shared encoder and shared post-training objective. Instead of separate vision/audio towers and task-specific outputs, it uses one shared encoder with unified space-time positions plus one schema for 10 parsing tasks; training is unified by self-distilled on-policy token supervision rather than hand-designed rewards.
Key results
- 96.96 on OmniDocBench v1.6, ahead of TeleOCR at 96.91
- 75.08 on OmniParsingBench, beating Logics-Parsing-Omni's 72.26
- 54.2 [email protected] on chemistry vs DeepSeek-OCR 2 at 49.6
A mostly shared encoder plus a unified output schema can cover broad multimodal parsing without giving up specialist-level precision.