OpenDecision-Large
Tokz Labs' open decision encoders on DeBERTa-v3. Choice, yes/no, multi-label and score questions over options defined at call time, several per call, with a probability per option and no generated text. Apache-2.0 weights plus a hosted Decisions API.
Trained on fastino/fast-decisions plus synthetic decisions aimed at its schemas, and public datasets listed with their licences in the technical report (sources that do not allow commercial use were removed). The policy head is not calibrated as shipped: top-option ECE is 22.94 points, 4.41 after a fitted temperature (the outcome head's ECE is 6.44). On the maker's classification suite (AG News, CLINC150, IMDb, Rotten Tomatoes, XNLI) Large averages 87.91 macro-F1; XNLI is not zero-shot for these models, and the suite guided the choice of training mixtures. Held-out label sets (TREC, Emotion, Subjectivity) average 57.00. Siblings: OpenDecision-Large-Packed (same size, up to 16 options per pass, order-invariant; 66 ms batch-1 p50 on an L4 for a short text with 10 options) and OpenDecision-Small (70.6M on DeBERTa-v3-xsmall, CPU-friendly). English. Hosted at $0.035 per million input tokens for Large and $0.025 for Packed and Small, answers free.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 434M |
| Base model | microsoft/deberta-v3-large |
| Maker | Tokz Labs |
| Released | 2026-10-04 |
| License | apache-2.0 |
Hosted API
Served by Tokz Labs — $0.035/MTok input, $0/MTok output. Get access · API docs.
Get the weights
pip install systemonemodels
systemone pull tokz-labs/opendecision
The files are served from the maker's Hugging Face repository, , and verified against the checksums recorded here.