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

- Fastino dropped GLiNER2.5 and somehow made entity extraction interesting. They killed span enumeration — the old approach of scoring every possible chunk width like a bureaucrat stamping every form in triplicate — and replaced it with boundary prediction. Start here, end there, done. Three Apache 2.0 checkpoints at 74M, 194M, and 287M params, all CPU-runnable. A two-person team with no GPU budget can now extract entities from 4,096-word documents without lighting a wallet on fire. Overall macro F1 hits 56.17 on 16 zero-shot benchmarks, with a 24.75-point jump on XNLI. Joint entity-relation decoding ships natively, so your output graph is schema-valid by construction instead of needing a post-hoc validation layer that nobody asked for. No inference provider hosts the checkpoints yet, so self-hosting is the only path — which is either freedom or a deployment headache depending on your stack-eye. Where's my cut? The cut is in the compute savings. This is the rare release that shrinks the model, expands the context, and doesn't promise to replace your entire engineering team. Aggressive passive income for anyone