September 2026 Summaries

Ollama 0.35 introduces support for decision models through TypeSafe’s Jev API, enabling locally run models to answer multiple typed questions about supplied text in a single request via the `/v1/systemone` endpoint. Designed for low-latency tasks such as ticket triage, model routing, and content or safety moderation, the feature avoids network overhead and demonstrated an average decision time of 91 milliseconds for Nimble 9B on an M5 Max MacBook Pro in a Pac-Man example. Three models are initially available: Bespoke Labs’ open-source 9B-parameter Nimble and Together AI’s experimental Tev1 models in 4B and 0.8B sizes. Requests can define choice, binary-like “noul,” and scoring questions, with responses providing selected answers, probabilities, confidence values, and token usage. Ollama plans further support for additional decision models, cloud-served models, and improved Apple Silicon performance through MLX.