Fastino Releases GLiNER2.5: A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction

Fastino has pushed out GLiNER2.5, a new take on information extraction that swaps out span enumeration for boundary prediction. That sounds technical, but the practical move is simple: instead of checking every possible span against a width grid, the model learns where an entity starts and ends.

That shift removes the old maximum-entity-width cap and opens the door to 4,096-word context windows. It also keeps compute linear in sequence length for a fixed schema, which is the sort of sentence that makes infrastructure teams sit up a little straighter. The model can now do joint entity and relation decoding, enforce cross-task label constraints, and attach attributes to spans in one pass.

Fastino is shipping three checkpoints on Hugging Face under Apache 2.0: 74M, 194M, and 287M parameters. The release notes say the 74M and 194M models can run on standard CPU boxes, while local inference is available on CPU, CUDA, or MPS through pip install "gliner2[local]" with Python 3.10 or newer. There’s no hosted inference provider for the checkpoints yet, so if you want to use them, you self-host.

On the benchmark side, Fastino reports results on 16 zero-shot datasets. The multilingual checkpoint posts an overall macro F1 of 56.17, just ahead of GLiNER2’s 56.09, while the base checkpoint reaches 54.87 versus 53.34 for GLiNER2. The biggest headline is XNLI: the multilingual model jumps to 62.30 from 37.55, a gain of 24.75 points.

The pitch here is pretty clear. Smaller teams get a model family that can run locally without GPU drama, and bigger teams get a privately hosted alternative to per-token LLM extraction. Fastino is aiming at legal work, clinical notes, insurance claims, customer support, guardrails, and anything else where entities, relations, and attributes need to come out cleanly instead of becoming another prompt-engineering hobby project.