Fastino Releases Open-Source GLiNER2.5 Models

Information extraction underpins systems that turn contracts, reports and customer records into structured data, but conventional approaches often enumerate possible text spans, increasing the candidate set and computing burden as documents grow. Fastino has positioned GLiNER as a compact alternative to large language models for named-entity recognition, classification and structured extraction. Local execution also gives companies a way to keep sensitive material on their own infrastructure rather than sending it to an external API.

Fastino released GLiNER2.5 on Aug. 25, 2026, replacing fixed-width span enumeration with sparse start-and-end boundary pairing. The company says inference now scales linearly with document length, supports contexts of up to 4,096 words and can capture spans of any length within the encoded window. Three Apache 2.0 checkpoints are available on Hugging Face: a 74 million-parameter small model, a 194 million-parameter English base model and a 287 million-parameter multilingual version. The models can run locally on CPUs, Nvidia CUDA hardware or Apple MPS devices.