On September 23, 2026, Fastino released GLiNER2.5-Decide on Hugging Face: a 340M-parameter open-weights encoder built for production decision work — intent detection, sentiment, topic classification, ticket routing, moderation, and spam filtering. It is deliberately not a general assistant. The model card states it does not reason, explain, or answer open questions. It is a specialist for operational decisions.

One model replaces a rack of classifiers

The traditional setup runs three separate classifiers for support intent, urgency, and routing — or prompts a general-purpose LLM and pays generative latency on every call. Decide takes a different route: the label set is a runtime argument. The same text and candidate labels go through a single forward pass; single-label tasks return one string, multi-label tasks return every label above threshold. The model card's examples cover banking intent, travel rebooking, clinic triage, aspect-level review sentiment, document type, ticket routing, human-handoff gating, moderation, and incident severity rating.

The most practical point for deployment: no prompt template, no generated tokens. A generative LLM has to compose an answer for every classification call; an encoder emits structured output directly, landing latency and cost in a different class. The gliner2 library installs with one pip line, runs local inference without a GPU, and the project positions itself as CPU-first with fully local processing.

340M vs 4B: what the benchmark shows, and its salt

The model card publishes the fast-decisions benchmark: 17 domains, 300 held-out examples per domain, identical text and candidate labels for every model. GLiNER2.5-Decide averages 60.2%, ahead of GLiNER2 XL (1B, 59.6%), JevK5 (57.6%), SemIf (a Qwen3.5-4B generative, 56.4%), GLiFormer large-v1 (49.0%), and Laya Router (46.6%). A 340M encoder outscoring a 4B generative baseline is the release's core selling point.

Before taking that at face value, note the salt: the table comes from Fastino's own dataset, with no third-party reproduction to check whether comparison models were tuned fairly or whether the label-set design favors the home architecture. Treat self-reported numbers with the usual discount. Notably, the lane is crowding: Convai's Laya Router, also pitched at routing and safety checks, sits last in the same table at 46.6%.

Boundary architecture and Chinese support

GLiNER2.5 uses a boundary architecture: sparse start-end pairing instead of a fixed-width span grid, so any span length inside the encoded window is representable. The GitHub repository (fastino-ai/GLiNER2, Apache 2.0, 1.8k stars) maintains both span and boundary architectures behind one public API, with AutoExtractor dispatching on the saved architecture field. Chinese users get a bonus: GLiNER2 supports a character-level word splitter, so Chinese does not fall back to the English default. The methodology has an EMNLP 2025 System Demonstrations paper behind it.

So what

The direction is clear: not every NLP task deserves a large model. High-frequency, low-complexity, latency-sensitive decisions run cheaper and more controllably on a 340M specialist on CPU than on a generative API billed per call. But a vendor-built benchmark is the vendor's own ruler — run your own label set through a held-out pass before adopting; that beats trusting any 60.2%. The model ships under a non-standard "other" license, so read the terms before commercial use.

References: fastino/GLiNER2.5-Decide · fast-decisions dataset · fastino-ai/GLiNER2