System One models read the state you provide and return typed decisions with probabilities, not generated prose. Your code keeps control of what happens next.
Models with published evidence of typed, non-generative decisions.
The first publicly announced System One model. Jev accepts program state plus typed questions and returns structured decisions with probabilities and confidence.
By TypeSafe AI · announced September 15, 2026ChoiceScoreNoulA 421M English decision model with a bidirectional encoder and option-scoring head. Returns typed probabilities without text generation. Apache 2.0; 512-token budget. Multilingual and workflow-specific checkpoints are separate variants.
By Convai InnovationsChoiceScoreNoulA Thai and English decision model built on Qwen3.5-0.8B with a dedicated slot head. Scores supplied choices, ordered levels, and yes/no questions without generating text. Apache 2.0; up to 255 choice options.
By iApp TechnologyChoiceScoreNoulAn English decision model using ModernBERT-base, a LoRA adapter, and a two-layer scoring head. Returns probabilities over typed options without text generation. Apache 2.0; 512-token cap. Published evaluations cover a limited set of classification tasks.
By Mateo LafalceChoiceScoreNoulInclusion describes the decision interface, not equal capability or calibration. Community models have not been benchmarked or integration-tested by HarnessRouter. Self-hosted inference has hardware costs; latency figures are not directly comparable.