GLiNER, GLiNER2, and prompted LLMs on 46 NER datasets — and the three ways a zero-shot NER leaderboard quietly misleads you (unfair aggregation, no uncertainty, inflated absolutes), plus how to fix each.

ONNX Runtime, batch inference, and tokenizer-based chunking for Presidio's NER recognizers, benchmarked across five hardware platforms. The winning configuration flips with hardware class.