Microsoft Unveils Microsoft-Decision-1 , a Decision AI Model Built on Alibaba 's Qwen3.5

Microsoft has launched Microsoft-Decision-1, a new artificial intelligence model that does not write paragraphs or draft emails. Instead, it reads an input and scores a fixed list of answer options, returning a confidence probability for each one ¹.

The model is built on top of Alibaba's open Qwen3.5-9B language model and is now available through Microsoft Foundry and OpenRouter.

What Makes It Different

Most AI models today are large language models, or LLMs, which generate text step by step. Microsoft-Decision-1 belongs to a newer category called decision models. These models take a question and a set of predefined choices, then output a structured probability score for each option in a single pass ².

This design makes them especially useful for tasks like routing support tickets to the right team, classifying content, prioritizing work items, or checking whether an AI agent's action passes a safety rule. The output is already in a format that software can act on immediately, without extra parsing ³.

Speed and Accuracy

In Microsoft's testing across 36 benchmarks containing roughly 147,000 unseen questions, Microsoft-Decision-1 achieved an average accuracy of 83.5 percent, leading all compared systems ². It also proved remarkably fast: the median response time was 85 milliseconds, about 2.5 times faster than the closest decision-model rival and roughly 35 times faster than large chat models like GPT-6 Sol ¹.

Internal tests show the model is resilient to small changes in wording or formatting, shifting its answer only about 1.3 percent of the time. Security testing across thousands of prompts found it blocked harmful requests while staying useful ¹.

Real Work Already Underway

Microsoft says the model is already being used inside the company. The Xbox Research team used it to label more than 10,000 reviews from surveys, Steam and X, matching the quality of GPT-6 Sol but at more than 14 times the speed and one two-hundredth of the cost. The Copilot team uses it to evaluate chatbot and agent responses, and Microsoft Discovery applies it to score experiments for adaptive planning ¹.

Microsoft plans to rebuild future versions on its own MAI models and on models from other providers such as OpenAI. The pricing is set at $0.042 per million input tokens, with outputs free ².

Key facts

• Built on Alibaba's open Qwen3.5-9B model
• Returns calibrated probabilities for fixed answer options instead of generating text
• 83.5% average accuracy across 36 benchmarks with 147,000+ unseen questions
• Median latency of 85 milliseconds, up to 35x faster than GPT-6 Sol
• Already used internally by Xbox, Copilot and Microsoft Discovery teams

Seen on social

Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident… pic.twitter.com/M3fAOD90r6

— Satya Nadella on X, see the post

Sources

1. Microsoft представила ИИ-модель на базе Qwen - ForkLog — forklog.com
2. Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision ... — marktechpost.com
3. Microsoft Decision-1: 9B Qwen Model Launch Explained — shattered.io
4. Alibaba Unveils Qwen3.8-Max: Its Largest and Most Capable Flagship ...
5. Microsoft Build 2026