Ollaya Runs Jev-Style Decision Models Locally
Ollaya's request ceiling of 255 options against Ollama's 15 widens what a local decision model can be asked to choose between, and the software ships under a permissive license rather than as a hosted service.
Key Facts
- Ollama added support for running Jev-style decision models in version 0.35, covering the models nimble, tev1, and tev1:0.8b.
- Ollama permits a maximum of 15 options per request, while Ollaya supports up to 255.
- Jev is an AI model that outputs judgments and confidence scores for multiple options, and it appeared on September 15, 2026.
- Ollaya supports Windows, macOS, and Linux, and its source code is published under Apache License Version 2.0.
- Ollaya ran the decision model Winnow-E4B locally on a PC with a GeForce RTX 4070 for an accuracy and latency comparison against Jev's API service.
Reporting from 1 source: GIGAZINE.
A new execution app called Ollaya runs Jev-style decision models on local hardware. Jev, which appeared on September 15, 2026, is an AI model that outputs judgments and confidence scores for multiple options. Ollaya supports Windows, macOS, and Linux, and is published as open source under Apache License Version 2.0. In a listed example, the decision model Winnow-E4B reads a user message about a third double charge and returns Intent: refund request (91%), Urgency: present (92%), and Anger level: 2.89 (86%). A chart comparing Jev's API service against a local run on a GeForce RTX 4070 shows that, depending on the model, local accuracy gets fairly close to Jev's while latency is shorter. Ollama added decision-model support in version 0.35, but that release covers three models, nimble, tev1, and tev1:0.8b, and allows a maximum of 15 options per request. Ollaya can run higher-performance models and set up to 255 options.
Ollaya's GitHub repository describes the software as "Ollama for decision models," a pull-and-serve tool that runs Laya, decider, NLI and GLiClass models behind a TypeSafe-compatible API.
Ollama 0.35 is the first entry in a run of releases built around decision models. Planned updates include MLX support to speed up operation on Mac and a wider set of supported models. Ollama also published footage of the nimble model playing a racing game by deciding to move left or right.
Ollaya lands in a crowded stretch for the format. OpenRouter's Jev Router uses Jev to pick the best AI model for a task, and on September 30, 2026 the decision model d1 appeared with benchmark scores above Jev. A free Jev-compatible model called Jeff runs locally in about 22 to 28 milliseconds. Jev itself beat Pokemon Red in 37 hours and 40 minutes for roughly 260 yen, though it needed a support system and could not leave Pallet Town when running fully autonomous.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.