GLiNER2.5 scraps span enumeration, hits 56.17 macro F1 across 16 benchmarks

Fastino just shipped GLiNER2.5, and the architectural change is the kind that quietly reshapes what’s practical for extraction teams. The old approach — enumerating every candidate span against a width grid — meant compute scaled with how long an entity could be. GLiNER2.5 throws that out. The model now predicts where entities start and end, scoring token boundaries instead of thousands of width combinations. That single decision removes the maximum entity width entirely, allows a 4,096-word context window, and keeps computation linear in sequence length for a fixed schema.

Across 16 zero-shot benchmarks, the multilingual checkpoint lands at 56.17 overall macro F1 versus 56.09 for its predecessor. The headline number is XNLI, where Multi jumps from 37.55 to 62.30 — a 24.75-point gain. Few-NERD improves for Base to 55.14 from 47.22, and Romanian RONEC, a language the model wasn’t trained on, improves for both sizes. Three checkpoints are live on Hugging Face under Apache 2.0 at 74M, 194M, and 287M parameters, all runnable on CPU.

What’s genuinely new here isn’t just speed. Joint extraction now decodes entities, typed relations, and structural rules in one beam search, so the output graph is schema-valid by construction — no post-hoc filtering needed. Constrained classification lets labels imply or exclude each other. Fastino’s own GLiGuard model illustrates why: without constraints, a prompt can be labeled safe while simultaneously flagged for prompt injection. With C.implies and C.excludes rules, that contradiction becomes impossible. If no valid assignment exists, the classifier raises an error rather than returning nonsense.

One catch: no inference provider hosts these checkpoints yet. Self-hosting is the deployment path, which is fine for a two-person team on a standard CPU box but adds friction for teams that wanted a managed API. For anyone doing PII redaction, contract clause extraction, or clinical entity work at volume, the economics shift meaningfully — this is the first time unlimited span length doesn’t cost extra compute.