Fastino Releases GLiNER2.5-Decide: A 340M Decision Model
Fastino launches GLiNER2.5-Decide, a 340M open-weight decision model running on CPU and GPU under Apache 2.0, excelling in classification and operational guardrails.
Stock photo for illustration only, not from the actual event
- Fastino released GLiNER2.5-Decide, a 340M parameter open-weight decision model running on CPU and GPU.
- Built as a non-generative classifier utilizing the DeBERTa-v3-large encoder architecture.
- Ranked first in 9 out of 17 evaluation datasets in Fast Decisions, excelling in intent routing.
- Available under the Apache 2.0 license and easily installable via pip install gliner2.
The Fastino team has officially launched GLiNER2.5-Decide, a 340-million parameter open-weight decision model designed to execute on central processing units, graphics cards, and air-gapped environments. Distributed under the Apache 2.0 license, the model can be installed seamlessly via pip install gliner2, with hosted API options and fine-tuning support readily available.
Architecturally, GLiNER2.5-Decide operates as a non-generative classifier fine-tuned from gliner2-large-v1 using a DeBERTa-v3-large encoder. The model generates no output tokens and requires no prompt templates. Label sets are supplied dynamically at runtime, where each schema question outlines permitted answers, cardinality expectations, custom instructions, examples, and cross-question validation rules.
Stock photo for illustration only, not from the actual event
The operational pipeline consists of two primary stages. First, the encoder processes text and schema simultaneously to evaluate scores for every allowed answer. Subsequently, a constrained decoder searches for the optimal joint assignment that satisfies all declared rules. Fastino explicitly notes that the model is specialized solely for operational decisions rather than open-ended reasoning, explanation, or conversational question-answering.
To demonstrate practical utility, Fastino highlighted a guardrail use case. When evaluating prompt injection attempts, joint decoding applies strict validation rules—forcing an unsafe verdict if any indicator of harm is detected. This returns synchronized safety and harm classifications that downstream software applications can directly utilize to block, route, or escalate incoming requests.
Employing a non-generative classification framework rather than a traditional generative large language model offers distinct advantages for operational workflows. By eliminating token generation entirely, the system mitigates hallucination risks and provides deterministic structured outputs that integrate reliably with downstream automation pipelines.
Fastino benchmarked the model against Fast Decisions, an internally curated held-out evaluation suite comprising 5,100 test examples distributed across 17 diverse datasets. The benchmarks spanned customer operations, domain routing across banking, clinical, travel, and benefits sectors, and general content comprehension, evaluated using exact-match accuracy.
GLiNER2.5-Decide secured the top position across 9 of the 17 benchmark datasets, demonstrating exceptional proficiency in intent routing. It achieved scores of 75.3% on support intent and 64.3% on banking intent, outperforming competing models by substantial margins. Additionally, Fastino introduced variant checkpoints including the 1B parameter GLiNER2.5-Decide-1B and the 287M multilingual GLiNER2.5-multi-Decide model.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment