Fastino Releases GLiNER2.5-Decide Open-Weight Decision Model

Fastino Labs has introduced GLiNER2.5-Decide, a 340 million-parameter open-weight decision model designed to streamline operational judgments within AI agent pipelines. This non-generative classifier accepts text inputs alongside a schema of typed questions and provides structured answers, each accompanied by a probability distribution, a confidence score, and metadata indicating constraint feasibility. The model is specifically engineered for frequent decision-making tasks such as routing, triage, tool selection, and implementing guardrails. The weights for GLiNER2.5-Decide are distributed under the Apache 2.0 license, allowing for straightforward installation via `pip install gliner2`. A key feature is its ability to run on standard CPUs, making it accessible even in air-gapped environments or on hardware without dedicated GPUs. Fastino Labs also offers managed inference and fine-tuning services through its GLiNER API for users who prefer a hosted solution.

GLiNER2.5-Decide operates as a classifier rather than a generative model. It utilizes a DeBERTa-v3-large encoder and has been fine-tuned from the `gliner2-large-v1` model. Unlike generative models, it does not produce any generated tokens and does not require prompt templates. Instead, label sets are provided at the time of the function call. Each question within the schema specifies its allowed answers and indicates whether it expects a single answer, multiple answers, or an ordered value. The schemas can be further enriched with instructions, examples, descriptions of labels, and rules that govern the relationships between answers across different questions. The model's processing pipeline consists of two stages: first, the encoder processes the text and schema together to score all permitted answers; second, a constrained decoder searches for the highest-scoring joint assignment that adheres to the declared rules. Fastino emphasizes that GLiNER2.5-Decide is a specialist tool for operational decisions and does not perform reasoning, explanation, or answer open-ended questions.

The significance of joint decoding is highlighted by Fastino through a guardrail example. When decoded independently, the model might flag a prompt injection attempt with a score of 0.82 and simultaneously label the same prompt as safe with a score of 0.52, leading to conflicting outputs. Joint decoding resolves this by applying a rule that mandates an unsafe verdict if any detected harm is present. Consequently, the model would consistently return `safety=unsafe` and `harm_type=prompt_injection` together. This unified output allows downstream code to reliably block, route, or escalate the request. The schema mechanism enables the expression of complex implications, ensuring that decisions are made holistically based on all relevant constraints and rules. This approach enhances the reliability and interpretability of AI agent decision-making processes.