Microsoft-Decision-1 is a specialized decision-scoring model designed to deliver calibrated probabilities for structured classification and routing.
What the product does and how it is positioned
Microsoft-Decision-1 is an AI decision model designed for low-latency classification, verification, and workflow routing. Built by post-training Qwen3.5-9B, the system focuses on returning structured probability scores across fixed candidate options rather than generating long-form prose.
The model operates via a structured API supporting multiple decision formats, including binary questions, multi-option selection, numerical ratings, and rubric-based evaluations for autonomous agents and automated pipelines.
Source-supported ways to use the product
Evaluates candidate steps in an agentic workflow to choose whether an agent should proceed, halt, retry, or hand off to a human.
Sorts unstructured customer feedback, survey responses, and social posts into predefined analytical categories and themes.
Classifies incoming requests or system incidents by priority and type to direct tasks to appropriate operational pipelines.
Microsoft-Decision-1 is built upon a post-trained Qwen3.5-9B foundation optimized for fast, single-pass probability generation. Instead of generating conversational text, it evaluates an input against a predefined set of potential outcomes and yields calibrated confidence values for each candidate.
Because the probability values are intended for direct programmatic consumption, the model is tuned to maintain probability calibration so that output confidence corresponds to empirical accuracy across varied distribution benchmarks.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
The model provides calibrated probability scores across fixed answer sets to support structured classification, routing, and scoring.
Unlike generative language models designed to produce text, this model is purpose-built to return structured outputs for automated software decisions.
The model accepts binary yes/no choices, multiple-choice options, rating scales, and rubric-based grading criteria via its API.
The initial model is post-trained on Qwen3.5-9B, with future versions planned for other architectures including Microsoft AI and OpenAI models.
The model is available for access through Microsoft Foundry and OpenRouter.
