A decision model that fits in a phone, already ported to one. GLiNER2.5-Decide is Fastino’s 340M-parameter open-weight classifier: hand it text plus a typed schema of questions (intent, routing, sentiment, priority, multi-label tags, ordinal scores, a question over a passage) and it hands back structured answers with confidence scores in a single forward pass - no prompt template, no generated tokens. 29,199 downloads in its first week. Apache 2.0, GLiNER2 architecture, 1.95 GB in fp16.
Benchmarks, labeled. On Fastino’s own fast-decisions suite (17 domains, 300 held-out examples each) it scores 60.2% average exact-match, ahead of JevK5 (57.6%), SemIf at Qwen3.5-4B scale (56.4%), and Laya Router (46.6%). Treat the suite as vendor-run: it is their dataset, their split, their comparison. What the number does establish is the schema trick - one model answering whatever fixed question set you pass at call time, instead of one classifier per job.
What runs where. This is the CPU end of the decision-model map: 167 ms per call on a 48-core server CPU with no GPU at all, and a 1.95 GB fp16 footprint. The port ecosystem moved fast - an official LiteRT build landed two days after release for phone GPUs, and a community CoreML port exists for Apple devices. A multi sibling (287M) covers non-English input, and a 1B variant trades a little memory for 59.6% on the same suite.
What it is not. Not a reasoner and not a writer: no explanations, no open questions, no generated text - the card itself says the model is a specialist for operational decisions, and outputs below a threshold get dropped rather than guessed. Anything needing judgment or synthesis stays with the reasoning model.
Where it sits. Jev made typed decisions credible at flagship scale, AutoJev reproduced it open at 27B, Jeff fine-tunes it home at 2B, and GLiNER2.5-Decide is the fixed-schema industrial end - a model that never writes, answers exactly the questions you give it, runs on the hardware a helpdesk already owns, and costs the electricity of a lightbulb.