Published: 25 September 2026
Fastino released GLiNER2.5-Decide on 24 September 2026, a 340M-parameter open-weight model built to classify and route text rather than talk. It scores 60.1% average accuracy on Fastino’s own 17-dataset Fast Decisions suite, ahead of JevK5 at 57.5%, and answers in 167 ms on a 48-core CPU with no graphics card at all. The licence is Apache 2.0.
What can GLiNER2.5-Decide do?
It answers a fixed set of questions about a piece of text, all at once. You hand it a support ticket and a list of typed questions: which product this is about, how urgent it is, whether it needs a human. The model returns structured answers in one forward pass, so one ticket costs one call instead of several.
That makes it a router, not an assistant. Fastino is direct about the limit. The model does not reason, explain, or answer open questions. What it does instead is decide, the same way every time, which is what most production classification actually needs.
You define the labels at the moment you call it, so adding a new category does not mean retraining anything. Alongside classification it handles entity extraction with character-level spans, relation extraction and structured record generation, with rules enforced across all of them.
The practical details:
- 340M parameters under Apache 2.0, so commercial use needs no negotiation
- Two siblings shipped the same day: GLiNER2.5-Decide-1B and GLiNER2.5-multi-Decide
- Install with pip install gliner2, then load it with AutoExtractor.from_pretrained
- Runs on CPU or GPU, with weights on Hugging Face at fastino/GLiNER2.5-Decide
- A hosted API at agent.fastino.ai, if you would rather not run it yourself
- Not a general-purpose model: no reasoning, no explanations, no open-ended answers
GLiNER2.5-Decide benchmarks and speed
Take this from the table: a 340M encoder beats a 4B language model at picking the right label, and it does it without a graphics card.
| What is measured | GLiNER2.5-Decide (340M) | The alternative |
|---|---|---|
| Fast Decisions average, 17 classification sets | 60.1% | JevK5: 57.5% |
| Same suite, against a 4B language model | 60.1% | SemIf (Qwen3.5-4B): 56.4% |
| Same suite, against a same-size encoder | 60.1% | Laya (ModernBERT, 421M): 46.6% |
| Support intent routing, single task | 75.3% | not published for rivals |
| Answer time, NVIDIA T4, 64 tokens | 43.6 ms | A100: 47.3 ms |
| Answer time, 48-core Xeon CPU | 167.3 ms | no GPU needed |
These are Fastino’s numbers on Fastino’s test set: 5,100 examples across 17 domains, and the company says plainly that this is an internal benchmark and not JevBench. It does not say where the figures for JevK5, SemIf and Laya came from, so read the gaps as the vendor’s framing rather than an independent result. The response times are median, at batch size 1.
What Fastino is not saying
There is no score on a public benchmark that anyone else runs. That matters more here than usual, because the entire claim is a comparison against named rivals, and Fastino picked both the tests and the opponents. A 3.6 point lead on a private suite is a reason to run your own evaluation, not a reason to skip one.
Throughput is missing too. Every speed figure is batch size 1, which tells you how fast one ticket comes back and nothing about how many you can push through in an hour. For the workload this model is aimed at, that second number is the one that sets your bill. Memory footprint is not published either.
What this means
Worth testing now if you are paying a frontier model to do clerical sorting. The concrete case is a four-person support-tools team putting a few hundred thousand tickets a month through a GPT-6 Luna call each, purely to decide where they go. At 340M parameters this runs on the web servers you already rent, the answer comes back in roughly 167 ms on plain CPU, and Apache 2.0 means nobody has to approve anything. Push a week of your real tickets through both and compare how often the labels agree. That test costs a day, and if it holds, the saving is permanent.
Safe to ignore if you need the model to say why. There is no explanation, no reasoning trace and no graceful answer on an input your labels do not cover, and Fastino says so itself rather than burying it. The wider pattern is worth noting. This is the third small specialist model in a month sold on being cheap and narrow rather than capable and general. The frontier labs compete on what a model can do; a growing tier underneath them competes on what you no longer have to pay them for.
For more information, visit the official announcement of GLiNER2.5-Decide on the Fastino blog.
