DharmaOCR tops OCR bench, cuts cost

This is a strong reminder that specialization can beat bigger general-purpose models when the task has a rigid output format and a measurable failure mode.

• –DPO here is not just alignment theater; using degenerate generations as rejected samples directly targets the looping and runaway outputs that hurt OCR pipelines.
• –The reported scores, 0.925 for the 7B model and 0.911 for the 3B model, are impressive, but they are still benchmark-specific, so the claim is strongest for structured OCR rather than broad document understanding.
• –AWQ cutting per-page cost by about 22% with negligible quality loss is the part that matters operationally, because OCR workloads are usually judged on throughput and unit economics as much as accuracy.
• –The comparison set is broad, spanning commercial APIs and open-source OCR stacks, which makes the result more interesting than a narrow internal eval.