Skip to main content

Decision Models

A decision model answers questions about an input with probabilities instead of writing text. You give it a state (a message, a JSON object, or an image) and questions, and every answer comes back typed:

Question typeWhat you defineWhat comes back
Yes/noA question, optionally what yes and no meanThe probability that the answer is yes
Choice2 to 255 named optionsThe likeliest option and every option's probability
Score2 to 10 levels, lowest firstThe probability-weighted level, counted from 1 in Chat and from 0 in the API

They are fast and cheap, which makes them good at routing, classification, moderation, and checking a draft against a rubric. They cannot write.

Where decision models come from​

ProviderHow Libre WebUI finds themAPI
OllamaInstalled models with the decision capabilityPOST /v1/systemone
llama.cppModels whose listing reports the decisions output modalityPOST /v1/systemone
BasetenThe decision-models/… namespace of its model listPOST /v1/systemone
OpenRouterIts catalog of decision modelsPOST /api/alpha/decisions
OpenAIgpt-6-luna-decisions, sent to OpenAI as gpt-6-lunaPOST /v1/decisions
Geminigemini-decision-1POST /v1beta/decisions

Ollama added decision models in 0.35.0, and newer models need newer releases (Clef Flash 0.35.1, Kev 0.40.4). When a pull says the model needs a newer Ollama, update Ollama and pull again. Pull a decision model from the model picker's Ollama tab under Decision, or from a terminal:

ollama pull nimble

Nimble and Tev read text; Clef and Clef Flash also read images. Provider decision models use the API key you already saved for that provider. On OpenAI, gpt-6-luna stays a chat model; its decision model is listed as gpt-6-luna-decisions so the two never mix. A llama.cpp server or an OpenAI-compatible endpoint with a changed Base URL sends its decisions to the same place.

In Chat​

Decision models appear in the model picker with a Decision badge. Pick one and the composer shows its decision questions. There are two ways to use it.

With questions set. Click Set questions and add yes/no, choice, or score questions. Each message you send is then a new state, and the reply shows each question with its answer, that answer's probability, and the full spread. Attach an image to judge it with a model that reads images.

Without questions. The chat works like ollama run with a decision model: your first message is the state, and each line you send after it is a yes/no question about it, so you can paste several at once. Send /clear to start over with a new state.

The questions are saved with the chat. Web search, tools, thinking, and model comparison are hidden while a decision model is selected, because none of them mean anything to it. Titles, follow-up suggestions, and context compaction never call a decision model.

Decisions API​

POST /api/decisions answers one request. API tokens need the chat scope.

curl https://your-instance/api/decisions \
-H "Authorization: Bearer lwk_..." \
-H "Content-Type: application/json" \
-d '{
"model": "nimble",
"state": "I was charged twice. Please refund the extra payment.",
"questions": {
"refund": { "type": "noul", "instructions": "Is a refund requested?" },
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": { "billing": "Payments and refunds", "technical": null }
}
}
}'

Add providerType and providerId to pick a provider's model, for example "providerType": "plugin", "providerId": "baseten". images takes up to four PNG, JPEG, or WebP images as base64 or data URLs. Libre WebUI translates the request to each provider's format and returns the same answer shape from all of them, with the token usage the provider reported. GET /api/decisions/models lists the decision models your account can reach.

Access​

Decision models are for administrators until an administrator opens them to everyone with the Decision models card in User Management. Set DECISION_MODELS_ACCESS_MODE to admins or all-users to pin the choice and lock the card. The check runs on every chat turn and every API request. Provider decision calls appear on the Provider Usage page with their model and token counts.

Limits​

A request takes at most 64 questions and 4 images of 4 MiB each (8 MiB in all), which is the tightest limit any supported provider documents. Each provider can be stricter: Nimble takes at most 26 options in a choice, Baseten's models take 16,384 tokens of input, and Ollama's local server takes 64 KiB of text. A provider's own error is shown as it gave it.