Fastino open-sources GLiNER2.5-Decide decision model
Fastino has released GLiNER2.5-Decide, a 340M-parameter Apache 2.0 encoder for schema-defined decisions that it says beats 4B-class decoder baselines on its own classification benchmark while running on CPUs.
Fastino announced GLiNER2.5-Decide, an open-weight encoder for structured decisions such as routing, triage and classification. The model takes text plus a schema of typed questions, each with its allowed answers. It returns the chosen answers along with probabilities, confidence scores and metadata showing whether the full set of answers satisfies the schema's rules.
A constrained decoder searches for the highest-scoring joint assignment allowed by rules across questions. In Fastino's example, decoding questions independently let the model flag a prompt injection while also labelling the same prompt "safe". Joint decoding, under a rule that any detected harm requires an unsafe verdict, returned a consistent answer. The model can also extract spans and relations, supports full or LoRA fine-tuning, and runs offline on CPUs.
On Fast Decisions, an internally generated suite of 5,100 test examples across 17 datasets, the model averaged 60.1% exact-match accuracy and led on 9 datasets. The other models scored lower: JevK5 at 57.5%, SemIf at 56.4% (both built on Qwen3.5-4B decoders), GLiFormer at 49.0% and the 421M ModernBERT-based Laya at 46.6%. The largest margins came on support intent (75.3%, 18.6 points ahead) and banking intent (64.3%, 8.6 points ahead). Fastino notes the benchmark is its own and not an independent evaluation. It also says JevK5 is an open reproduction, not TypeSafe's Jev.
Reported p50 latency at batch 1 and 64 tokens was 38.3 ms on a V100, 43.4 ms on an L4, 47.3 ms on an A100 and 167.3 ms on a 48-vCPU Xeon. At 1,024 tokens the A100 pulled ahead, at 52.6 ms.