Ollama adds decision models for local classification and scoring

Ollama began supporting decision models with its 0.35 pre-release on September 28, 2026. These models do not write long text; they answer a question with a choice, a probability or a score. Ollama's example uses are routing support tickets to the right team, sending a request to the right model and classifying content.

How a decision model works

The new /v1/systemone endpoint follows the Jev API from a company called TypeSafe. You send the text to be judged and up to 64 questions. There are three question types: choice picks one option and gives each a probability, noul returns the probability that a condition is true, and score gives a score against ordered criteria. In the release notes' example, the text "Our checkout has returned 500 errors since 9am" is sorted into bug, rather than billing or account, with high probability. Each answer also comes with a confidence value; the model pages note that it is not the chance that the answer is right.

The first two models

Nimble is a 9-billion-parameter model that Bespoke Labs trained on top of Qwen3.5-9B. It takes about 9.5 GB, is licensed under Apache 2.0 and installs with ollama pull nimble. It is fast because it scores the answer tokens directly with no reasoning step; according to its page, the text being judged must fit within 8,192 tokens. Tev1 is an experimental model family from Together AI: the 4-billion-parameter version is 4.5 GB and the 0.8-billion version is 812 MB. Together AI shares its data builders and training scripts under the MIT license but does not state a license for the model on the page.

For companies running AI on their own servers, this is a simple way to sort support tickets and emails without sending data outside. From Python it can be used through TypeSafe's SDK, pointed at the local Ollama address.

Version 0.35 is still a pre-release; the stable version remains 0.34.4. Decision models are not yet in Ollama's command line or its Python and JavaScript libraries, and the announcement post has not been published yet. The models work only with text and only choose among the options you provide. Together AI writes that Tev1 has not been fully tested on languages other than English or against prompt injection, so it is worth trying it on your own examples first. To see how the models do on different tasks, it is worth following comparisons from independent researchers.

Sources: GitHub: Ollama v0.35.0 release notes, Ollama: Nimble, Ollama: Tev1.