Today, we release GLiNER-2.5-Decide, a 340M-parameter open weight model for schema-defined decision-making.

Given text and a set of typed questions, it returns valid answers, probabilities, confidence scores, and constraint-feasibility metadata. Alongside constrained decisions, GLiNER2.5-Decide can extract spans and relations, enforce rules across related outputs, and support full or LoRA-based fine-tuning. It runs locally on CPUs and can be deployed in air-gapped environments under the Apache 2.0 license.

Due to its deliberately compact architecture, GLiNER2.5-Decision is exceedingly fast. On short-document tasks, we report end-to-end p50 latency of 38.3 ms on an NVIDIA V100 and 167.3 ms on a 48-vCPU Intel Xeon Platinum 8581C.

Across our unseen 17-dataset benchmark, spanning classification, routing, triage, and content-understanding tasks, GLiNER-2.5-Decide achieves the highest average score at 60.1%. It outperforms SemIf, an established decision-model baseline built on Qwen3.5-4B, and Laya, a 421M ModernBERT encoder and the top-trending model on the Hugging Face Hub as of this writing.

These results show that a specialized, non-generative encoder can outperform much larger decoder readouts while also leading comparable encoder systems on schema-driven classification.

Benchmark Design

We evaluate GLiNER2.5-Decide on an Fast Decisions, an unseen, internally generated classification suite containing 5,100 test examples across 17 datasets. Every model receives the same state, question, and permitted answer set. We measure micro-task accuracy: a decision is counted as correct only if its predicted label set exactly matches the reference answer.

The benchmark spans:

• Customer operations: support intent, support topic, agent handoff, email triage, ticket routing, and product feedback.

• Domain-specific routing: banking, clinical requests, travel, and benefits.

• General content understanding: document type, review sentiment, news topics, academic fields, sports recaps, restaurant reviews, and screen tags.

We chose these tasks because they resemble decisions consumed directly by software, including routing requests, selecting queues, assigning topics, determining sentiment, or triggering a handoff. The mix tests performance across changing schemas, specialized vocabulary, different label inventories, and closely related choices.

Results and interpretation

GLiNER2.5-Decide achieves the highest average score at 60.1% and led 9 of 17 datasets. Particularly strong on intent routing, it reaches 75.3% on support intent and 64.3% on banking intent, 18.6 and 8.6 points ahead of the next-best models, respectively.

Across the full suite, GLiNER2.5-Decide outperforms JevK5 at 57.5%, SemIf at 56.4%, GLiFormer at 49.0%, and Laya at 46.6%. JevK5 and SemIf use 4B-class Qwen3.5 decoder models, while GLiFormer and Laya are single-pass encoders.

The results suggest that a specialized, non-generative encoder can outperform much larger decoder readouts while also leading comparable encoder systems on schema-driven classification. This is an internal benchmark, not JevBench, and JevK5 is an open reproduction rather than TypeSafe’s Jev.

Latency

We also benchmarked the released 340M-parameter checkpoint end to end using a two-head, 15-label classification schema. At batch 1 and 64 tokens, p50 latency was 167.3 ms on a 48-vCPU Intel Xeon Platinum 8581C, 43.6 ms on an NVIDIA T4, 43.4 ms on an L4, 38.3 ms on a V100, and 47.3 ms on an A100.

Short requests are dominated by fixed preprocessing and kernel-launch overhead, which keeps the tested GPUs within 9 ms of one another. The A100 becomes more effective as inputs grow, reaching 52.6 ms p50 at 1,024 tokens, compared with 75.6 ms on V100 and 131.4 ms on L4.

For low-volume deployments, CPU-only inference remains practical at 167.3 ms p50 for a 64-token document. For short online requests, L4 and V100 offer strong latency without requiring an A100.

Model Architecture, Inputs, and Outputs

GLiNER2.5-Decide is an encoder-based decision model post-trained specifically for structured decision-making. It evaluates user-defined typed questions and can decode related answers jointly under explicit constraints, returning structured decisions with probabilities, confidence scores, and feasibility metadata.

GLiNER2.5-Decide takes text together with a schema of typed questions. Each question specifies its permitted answers and whether it expects one answer, multiple answers, or an ordered value. The schema can also include instructions, examples, answer descriptions, and rules connecting answers across questions. The encoder processes the text and schema together, producing a compatibility score for every permitted answer. A constrained decoder then searches for the highest-scoring joint assignment allowed by the declared rules. The result contains the selected answer and probability distribution for each question, plus metadata indicating whether the complete decision satisfies the schema.

To illustrate why joint decoding matters, consider GLiNER2.5-Decide evaluating a prompt for both overall safety and harm type. When these questions are decoded independently, the model can detect prompt injection with a score of 0.82 while separately labeling the same prompt safe with a score of 0.52. The model has recognized the attack, but the two decisions conflict. Joint decoding evaluates the answers together under a rule that any detected harm requires an unsafe verdict. It therefore returns the coherentcombination safety=unsafe and harm_type=prompt_injection, with scores that downstream code can use to block, route, or escalate the request.

Beyond typed decisions, GLiNER2.5-Decide’s multi-task encoder-based architecture provides several additional capabilities:

• Character-level spans for extracted information. When a schema asks for entities or relations, results can include the exact start and end offsets of the extracted text. Classification answers themselves do not return evidence spans.

• Relations and structured records in the same model. A schema can combine entities, relations, classifications, and structured records, with the requested tasks scored in one forward pass.

• Constraints across classification tasks. Schemas can express implications, exclusions, cardinality limits, and ordinal bounds. The decoder searches for the highest-scoring joint assignment permitted by those rules and reports whether the result is feasible.

Model Capabilities and Use Cases

Decision models are particularly useful when software must make several related judgments from the same text. GLiNER2.5-Decide combines typed decisions with explicit constraints, span extraction, relations, and structured records, supporting use cases that require deterministic, low-latency decision-making.

1. Model routing. Route requests by destination, complexity, or escalation level and return the selected route with probabilities and confidence scores.

2. Tool calling. Select from a permitted set of tools and capture arguments from the request as structured records.

3. Browser and computer use. Evaluate the current state to select the next action. Represent the action, target, and parameters as a structured record.

4. Guardrails. Evaluate safety, harm type, and escalation together, extracting harmful spans verbatim. Apply constraints when combinations of answers would be contradictory.

5. Context pruning. Decide which information in the context is relevant and return the exact character spans associated with that decision.

6. LLM-as-a-judge. Express rubric criteria as categorical or ordinal questions and return selected values with probability distributions and confidence scores.

7. Games and simulations. Evaluate game state to determine intent, action, or outcome. Represent actors and state information through entities, relations, and structured records.

Getting started

• Model: https://huggingface.co/fastino/GLiNER2.5-Decide

• Demo library: https://fastino.ai/models/demo-library

• Source and documentation: https://docs.fastino.ai

• Inference API: https://agent.fastino.ai

• Agent skill: https://docs.fastino.ai/concepts/gliner-agent-skill