What it is
GLiNER2.5-Decide is a classification model from Fastino, billed as 340 million parameters (its weights file holds about 486 million, per the Hugging Face listing), published on Hugging Face in September 2026 under Apache 2.0. It is built on a DeBERTa-v3-large encoder and the GLiNER2 schema-driven library. You pass text and one or more label sets at call time, for intent, routing, sentiment, priority, moderation or multi-label tags, and it scores them all in one forward pass.
There are three versions: the English model, a 287M multilingual model and a 1B model. All run on a CPU or GPU through the gliner2 Python package, and Fastino publishes a fine-tuning recipe that turns each decision into a JSONL training row.
It is not a full System One model in the Jev sense: questions are label sets rather than typed choice, score and yes/no objects, and ordinal scores are trained as plain classes. The colibri engine and other gateways wrap it behind the Jev request format.
Our review
GLiNER2.5-Decide is the practical way to add flexible labels to a pipeline that cannot afford a GPU, provided the labels are operational. Pass the label list for this request and the answer comes back in a few hundred milliseconds on a CPU; colibri measured 294 ms for one question and 897 ms for three on a desktop processor, using about 1.9GB of memory. Labels can carry descriptions, several heads are scored in one pass, and fine-tuning needs only a JSONL file.
Fastino's own benchmark puts it first: 60.2% exact-match accuracy across 17 English domains of its fast-decisions dataset, ahead of a JevK5 model at 57.6% and Laya's router at 46.6%. Its larger 1B sibling scores slightly lower, at 59.6%, so there is little reason to pay for the bigger model on these tasks.
The community Jev Decision Index 0.4.1, published on October 21, 2026, tells a different story. It gives GLiNER2.5-Decide a skill score of 10.1, 215th of 229 entries, against 60.2 for Jev 1.14 and 68.0 for Perplexity's 27B Decider. Laya, the other small encoder in this class, sits just below it at 9.2. The suite covers knowledge, reasoning, tool selection and retrieval that a small encoder was never built for, but the gap is too wide to call it a general decision model.
Ordinal scales are its subtle trap. The model card says scores like 0 to 10 are trained as plain classes, so the loss does not know 6 is closer to 7 than to 0, and Fastino itself advises measuring such heads with mean absolute error as well as accuracy.
Where it earns its place
A 40-agent support team on a self-built helpdesk
Route every incoming ticket to refunds, login, shipping or a human before an agent sees it
Run the classifier on a CPU in the existing ticket service and pass the queue list as labels; no GPU or API bill. Fine-tune on a few thousand historical tickets before trusting it on edge cases.
A trust-and-safety engineer at a forum with a million comments a day
Put a cheap moderation gate in front of a slower LLM reviewer
Score allow, spam, harassment and personal-data labels on every post on commodity servers, and only escalate uncertain ones. The saving is the LLM calls avoided; the risk is misses on subtle policy cases.
An operations manager at a mid-size clinic group
Sort patient emails into appointment, refill, results and referral queues
Labels with short descriptions keep a private taxonomy precise. The model is English only, so the multilingual 287M variant is the one to evaluate for non-English patients.
Before you commit
- Setup
- pip install gliner2, load fastino/GLiNER2.5-Decide and call classify_text with a label dictionary; the first answer takes minutes to reach on a laptop.
- Integrations
- gliner2 Python package on CPU or GPU; colibri's C engine and other gateways wrap it behind the Jev request format; Fastino Agent offers hosted fine-tuning as a separate product.
- Data and privacy
- Apache 2.0 weights run entirely on your hardware, so text is never sent to Fastino unless you use its fine-tuning agent.
- Cost at scale
- Self-hosted only: about 1.9GB of memory per process and roughly 0.3 seconds per question on a desktop CPU, so a few commodity servers cover millions of decisions a day with no per-call fee.
- Limits
- English only for this checkpoint; label sets rather than typed choice, score or yes/no objects; weak on knowledge, reasoning and tool-selection tasks in the Decision Index.
What's good
- Under 0.5B parameters; runs on a CPU in under 300 ms per question
- Any label set at call time, with optional descriptions
- Several decisions scored in one forward pass
- Apache 2.0, with a 287M multilingual variant
What's not
- Skill score of 10.1 on the Decision Index 0.4.1, 215th of 229
- Headline 60.2% figure is from Fastino's own English-only suite
- Ordinal labels trained as plain classes
- No reasoning or explanations, by the model card's own statement
Best for
Engineers who need cheap, CPU-only intent routing, moderation or spam decisions and can fine-tune on their own labels.
Alternatives
GLiNER2.5-Decide: common questions
- Is GLiNER2.5-Decide worth it?
- GLiNER2.5-Decide is a fast, honest label classifier that performs well on Fastino's tasks and far behind on everyone else's. We rate it 3.6 out of 5.
- How much does GLiNER2.5-Decide cost?
- Free weights and library under Apache 2.0. Fastino's hosted fine-tuning agent is a separate product.
- What are the downsides of GLiNER2.5-Decide?
- Skill score of 10.1 on the Decision Index 0.4.1, 215th of 229; headline 60.2% figure is from Fastino's own English-only suite; ordinal labels trained as plain classes; no reasoning or explanations, by the model card's own statement.
- Who is GLiNER2.5-Decide best for?
- Engineers who need cheap, CPU-only intent routing, moderation or spam decisions and can fine-tune on their own labels.
- How much does GLiNER2.5-Decide cost for a team or in production?
- Self-hosted only: about 1.9GB of memory per process and roughly 0.3 seconds per question on a desktop CPU, so a few commodity servers cover millions of decisions a day with no per-call fee.
- Is GLiNER2.5-Decide safe for company data?
- Apache 2.0 weights run entirely on your hardware, so text is never sent to Fastino unless you use its fine-tuning agent.
- What are the best GLiNER2.5-Decide alternatives?
- The closest alternatives we have reviewed are Laya, Jev, Perplexity Decider (pplx-decider), Clef and d3 (vLLM Semantic Router).
Answers checked Oct 23, 2026.