Ollama v0.35.0 Adds Decision Models API for Classification and Routing Tasks

Ollama Ships Decision Models API in v0.35.0

Ollama released version 0.35.0, adding a new API endpoint, /v1/systemone, that supports "decision models" — a class of models built to return structured outputs like choices, probabilities, and scores rather than free-form text. The feature is based on TypeSafe's Jev API, according to the release notes.

Two decision models are available at launch: Nimble, from Bespoke Labs, and Tev1, from Together AI. Both can be pulled through the standard Ollama CLI (ollama pull nimble) and queried through the new endpoint.

How It Works

Instead of prompting for generated text, developers send a state (context) and one or more questions to the endpoint. The API supports three question types:

• choice: selects one option from a set and returns a probability for each
• noul: returns the probability that a stated condition is true
• score: returns a score across an ordered set of criteria

In Ollama's example, a support ticket describing checkout errors is classified against three labels — billing, bug, and account — with the model returning a bug classification at 97.81% probability and an overall confidence score of 0.8906. The response also includes token usage: 174 input tokens and just 1 output token, reflecting the compact nature of decision-model responses compared to generative text output.

Ollama positions the feature for tasks like support ticket triage, model routing, and content classification — use cases where a probability distribution over a fixed set of outcomes is more useful than a generated sentence.

Other Changes in v0.35.0

The release also includes several fixes unrelated to decision models:

• The Settings panel now opens immediately, without waiting on model discovery to complete.
• A macOS bug where the update menu and icon failed to reflect an available update at startup has been fixed.
• Stalled MLX model downloads that previously hung indefinitely are now resolved.
• Requests using the deprecated typical_p parameter now log a warning instead of failing outright.

The full changelog is available on GitHub comparing v0.34.4 to v0.35.0.

What This Means

Decision models address a narrow but common production need: many LLM deployments don't actually need generated prose, they need a classification, a routing decision, or a confidence score. By exposing a dedicated endpoint and pulling in third-party models purpose-built for this (Nimble, Tev1), Ollama is positioning itself as infrastructure for structured-output workloads, not just chat and code generation. The low output-token counts in these responses (often just one token) also mean lower latency and cost for high-volume classification tasks compared to routing that work through a general-purpose chat model with JSON-mode prompting. Whether Jev's underlying approach outperforms standard LLM-based classification on accuracy or calibration remains unverified — Ollama's release notes make no benchmark claims for Nimble or Tev1, and no accuracy figures are provided against established classification baselines.