Agents make a lot of small decisions. Which team gets this ticket. Which tags apply. Is it urgent. Most teams answer them by prompting a large language model, which works, but every call decodes tokens nobody reads and the output still has to be parsed.
OpenDecision is a family of open encoder models built for these calls. You send a text and the questions you want answered. You get back the best option and a probability for every option.
What you can ask
- Choice (
choice): exactly one option, like a queue or an intent. - Tags (
multilabel): any number of tags. - Yes / no (
bool): yes or no. - Score (
score): a value on an integer scale.
Labels are plain text in the request, so a new label set needs no training run. One request can carry up to 64 texts and 16 questions.
curl https://api.tokz.si/v1/decide \
-H "Authorization: Bearer $TOKZ_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"inputs": "I was charged twice for my May invoice.",
"decisions": {
"queue": { "type": "choice", "options": ["billing", "technical", "sales"] },
"urgent": { "type": "bool" }
}
}'Three models
| Model | Params | Macro-F1 | GPU-s / 1k | Per 1M tokens |
|---|---|---|---|---|
| OpenDecision-Large | 434M | 87.9 | 77.4 | $0.035 |
| OpenDecision-Large-Packed | 434M | 87.2 | 69.1 | $0.025 |
| OpenDecision-Small | 71M | 79.3 | 17.5 | $0.025 |
The suite is the average macro-F1 over the full test sets of AG News, CLINC150, IMDb, Rotten Tomatoes and XNLI. GPU time is GPU-seconds per 1,000 decisions on one NVIDIA L4.
How it compares
We ran six open decision models through the same harness on identical inputs. OpenDecision-Large scored highest on the suite. SemIf (86.3) and JevK5 (86.3), both built on a 4B language model, came close and needed about 3.5× more GPU time.
There is a limit worth knowing. On label sets that no training data targeted, the 4B models and Laya generalize better than ours. If your labels are unusual, test on your own data first. The technical report has both results.
Open, and hosted
The weights are on Hugging Face under Apache-2.0, and every training dataset allows commercial use. The code, the evaluation harness and every prediction behind the tables are public too.
Run the models on your own GPU, or call ours and skip the serving. Same weights either way. New accounts start with $5 of free credit.