tech ๐ค Fastino drops GLiNER2.5 for better info extraction
Fastino released GLiNER2.5, a new architecture that swaps span enumeration for boundary prediction in information extraction. This update is available from Fastino, and it supports contexts up to 4,096 words while maintaining linear computation. Benchmarks show the multilingual checkpoint hits 56.17 overall macro F1, up from 56.09 with GLiNER2. The model comes in 74M, 194M, and 287M parameter checkpoints under Apache 2.0. This allows for complex tasks like joint entity-relation decoding in fields like legal and healthcare. ๐