Decision engine

Classify text without an LLM

Ask typed questions about text and get a probability for every option.

Try it in the playground Read the report

POST /v1/decide: endpoint
from $0.025: per 1M tokens
Apache-2.0: open weights, same as the API
4 question types: choice, tags, yes / no, score

Tokz API / POST /v1/decide · illustrative

Text “I was charged twice for my May invoice.”

Illustrative

Which team should handle this?

queue

billing (most likely)0.97
technical0.02
sales0.01

Is it urgent?

urgent

no (most likely)0.82
yes0.18

Request

POST /v1/decide

{
  "inputs": "I was charged twice for my May invoice.",
  "decisions": {
    "queue": {
      "type": "choice",
      "question": "Which team should handle this?",
      "options": ["billing", "technical", "sales"]
    },
    "urgent": { "type": "bool" }
  }
}

Raw response

200 · illustrative values

{
  "results": [
    {
      "queue": {
        "best": "billing",
        "policy": {
          "billing": 0.97,
          "technical": 0.02,
          "sales": 0.01
        }
      },
      "urgent": {
        "best": "no",
        "policy": {
          "yes": 0.18,
          "no": 0.82
        }
      }
    }
  ],
  "usage": {
    "billed_tokens": 46,
    "cost_usd": 0.000001656
  }
}

Figure 1. Illustrative: one request, two questions, example probabilities.

Four question types

Up to sixteen questions on up to sixty-four texts per request. No training run.

Choice · choice
Exactly one option. Routing, intent, sentiment.
Tags · multilabel
Any number, each with its own probability.
Yes / no · bool
One question, with how sure the model is.
Score · score
A value on your integer scale, with the full distribution.

Pick a model

On label sets they weren't trained on, JevK5, SemIf and Laya score higher. Test on your own data.

OpenDecision / Accuracy against GPU time

Suite macro-F1 against GPU-seconds per 1,000 decisions
System	Suite macro-F1	GPU-seconds per 1,000 decisions
OpenDecision-Small	79.26	17.5
Laya	81.47	43.5
OpenDecision-Large-Packed	87.23	69.1
OpenDecision-Large	87.91	77.4
GLiFormer	80.53	109.5
JevK5 (4B LLM)	86.27	267.4
SemIf (4B LLM)	86.32	275.0

Figure 2. Suite macro-F1 against GPU-seconds per 1,000 decisions on one NVIDIA L4, model loading excluded. Open models only; GLiNER2.5 has no speed measurement. Measured 2026-10-04 and 2026-10-05 · Methodology.

OpenDecision / models