Fastino Releases 340M GLiNER2.5-Decide Open-Weight Classifier

Fastino's 340M Apache 2.0 GLiNER2.5-Decide classifier scores about 60% exact match on a 17-domain suite for routing, triage, and guardrails on CPU.

Fastino Labs released GLiNER2.5-Decide, a 340-million-parameter Apache 2.0 English decision classifier that was trending on Hugging Face and is designed for CPU or air-gapped deployment. Built on a DeBERTa-v3-large encoder and fine-tuned from gliner2-large-v1, the non-generative model takes arbitrary labels at call time and returns structured scores with probabilities, confidence, and feasibility through two-stage constrained joint decoding, without prompts or generated tokens. Sources disagree on accuracy: one reports 60.2% average exact match on fastino/fast-decisions across 17 domains with 300 held-out examples each, while a later report reports 60.1% on an internal 17-dataset Fast Decisions suite and says it led 9 of 17 datasets. Both put the 1B sibling at 59.6%; the first also cites SemIf (Qwen3.5-4B) at 56.4%, and the second reports the 287M multilingual GLiNER2.5-multi-Decide at 56.7%. MarkTechPost adds p50 latency of 167.3 ms on CPU at 64 tokens and 38.3 ms on an NVIDIA V100, with uses such as routing, triage, moderation, and guardrails via gliner2 AutoExtractor.

• Fastino Labs released GLiNER2.5-Decide, a 340M non-generative English classifier under Apache 2.0 that was trending on Hugging Face and is meant to run on CPU or air-gapped systems.
• It uses a DeBERTa-v3-large encoder fine-tuned from gliner2-large-v1, scores caller-supplied labels in one forward pass with two-stage constrained joint decoding, and returns probabilities, confidence, and feasibility without generated…
• Exact-match figures disagree: 60.2% on fastino/fast-decisions (17 domains, 300 held-out examples each) versus 60.1% on an internal 17-dataset Fast Decisions suite, where it led 9 of 17 datasets.
• Both sources put GLiNER2.5-Decide-1B at 59.6%; one cites SemIf (Qwen3.5-4B) at 56.4%, and the other cites 287M GLiNER2.5-multi-Decide at 56.7%.
• Reported p50 latency is 167.3 ms on CPU at 64 tokens and 38.3 ms on an NVIDIA V100.
• Uses include intent, routing, sentiment, priority, moderation, spam, triage, and guardrails; it loads locally with gliner2 AutoExtractor.