News
Latest news and updates on local artificial intelligence
8 news items in Decision models
An 11.4 MB open-weight decision model: the System-1 for devices
October 9, 2026Hugging Face hosts ej, a small 11.4 MB calibrated Decision model with open weights: given a state and typed questions, it returns one probability distribution per question in a single pass, without decoding tokens. Designed to run on CPU with a 128 MB resident-memory peak.
The Jev architecture was open-sourced a year ago: model, dataset and paper included
October 7, 2026On LocalLLaMA, one author describes building the architecture behind the Jev decision models a year ago and open-sourcing it completely: model, dataset and paper. Backstory that accompanies the arrival of the System One format in llama.cpp and the opening of the decision-model ecosystem to the community.
llama.cpp adds decision models via the /v1/systemone endpoint
October 7, 2026The llama.cpp server now supports decision models: send a state (text, JSON or a screenshot) plus typed questions, and the model returns a probability for each option in a single forward pass, without generating text. Three question types (choice, score, noul) and six models from 144M (Julia-1) to 27B (OpenJev, Clef), answering in 3-43 ms on an RTX PRO 6000, Apache 2.0 except OpenJev (CC BY-NC).
Ordinal-scale bias in JEV-like direct-decision models
October 1, 2026Research paper on direct-decision models: JEV 1.13 and three open KEV models do not use the ordinal scale supplied by the user faithfully. On ANLI, JEV assigns 38.8% of all predictions to the top option regardless of content — more choices, fewer decisions.
Cloudflare releases clef: 27B decision model, Apache 2.0
October 1, 2026Cloudflare made clef public, a 27B model (base Qwen3.8-27B) that turns a state — text, JSON, images, or video — and a schema of typed questions into structured decisions: one probability per option in a single pass, with no text generation. Jev/SystemOne API compatible; a smaller, faster variant clef-flash is available too.
Ollama 0.35.0: Jev-style decision models via the /v1/systemone endpoint
September 30, 2026With Ollama 0.35 you can use decision models based on TypeSafe's Jev API: send text as a state with a set of named questions and a local model answers them all in one request, with a probability per option and no free-form text generation. Built for ticket triage, model routing, and content moderation.
Ollama 0.35 brings decision models: nimble, a 9B model for typed classification
September 30, 2026Ollama 0.35 adds decision-model support: Bespoke Labs' 9B nimble model runs entirely locally and answers typed questions (true/false, numeric scale, multiple choice) through the new systemone API.
Jev in the Wild: data-driven analysis of Jev decision models in the wild
September 28, 2026CUHK study of Jev decision models: 2,050 real-world use cases analyzed, from triage and routing to moderation and labeling. The paper maps how Jev models work, their applications, and the ecosystem of models that answer with structured labels instead of free text.