Urchade Zaratiana has released GLiNER, an NER model created “that can identify any type of entity using a bidirectional transformer encoder”. One of the ToDos listed is „Allow longer context“, and a closer look reveals that:

• The Colab caps out at 384 tokens, and
• confirms DebertaV2 architecture.

Interesting, but too restricted as of now.