Ollama introduces support for Jev-style decision models for local AI processing

Ollama has announced support for decision models built on TypeSafe's Jev API, marking a significant step toward faster, local AI decision-making. The new functionality, available starting with Ollama 0.35, exposes a /v1/systemone endpoint that allows users to send text state along with a set of named questions, receiving typed answers in a single request. This enables rapid classification and decision tasks, such as customer ticket triage, model routing, and content or safety moderation, to run directly on local machines without network latency.

The performance gains are notable: Ollama's benchmark testing showed the nimble 9B model averaging just 91 milliseconds per decision when running locally on an M5 Max processor. The platform currently offers three new decision models, nimble, tev1, and tev1:0.8b, each designed for different computational constraints. Response payloads include probabilities and confidence scores alongside the primary decision output, giving developers visibility into model certainty. This transparency is particularly valuable for production systems where understanding model confidence can inform downstream handling.

Developers can integrate these decision models using Ollama's curl interface or TypeSafe's official Python SDK. Setup requires downloading the latest Ollama version, pulling a decision model like nimble, and configuring environment variables for the TypeSafe SDK. While the initial release focuses on local execution, Ollama indicated that additional models served through its cloud infrastructure are planned for future release.