Now on the Tokz API
OpenDecision
Open models that route, tag and score text, with a probability for every option.
Small models for AI agents
Open models that answer an agent's classification calls without an LLM.
- 87.9
- macro-F1, OpenDecision-Large, 5-dataset suite
- 3.6×
- less GPU time than a 4B LLM-based classifier
- 17.5
- GPU-seconds per 1,000 decisions, OpenDecision-Small
OpenDecision / Classification suite
OpenDecision-Large leads the open models we measured
Average macro-F1 on five classification benchmarks, full test sets, same harness.
Methodology in the reportModelMacro-F1
- OpenDecision-Large87.9
- OpenDecision-Large-Packed87.2
- SemIf (4B LLM)86.3
- JevK5 (4B LLM)86.3
- Laya81.5
- GLiFormer80.5
- OpenDecision-Small79.3
- GLiNER2.574.4
Open models only. On label sets they weren't trained on, JevK5, SemIf and Laya score higher.
Tokz API / Decisions
Decisions on the Tokz API
Put your labels in the request. No training run and no LLM.
POST /v1/decide
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"]
}
}
}'From the lab
Latest models
- 01OpenDecision-Large-PackedNewThe default: near-Large accuracy at the Small price, up to 16 options.In the API
- 02OpenDecision-LargeOpen weights434M parameters and the family's best suite score, 87.9 macro-F1.In the API
- 03OpenDecision-SmallOpen weights71M parameters for high-volume routing, at about a quarter of Large's GPU time.In the API
- 04tokz CLIOpen sourceShows where your coding agents' tokens and money went.Install
