Microsoft Launches Microsoft-Decision-1: A Post-Trained Qwen3.5-9B Model for Ultra-Low Latency Structured Decision MakingIT之家 · 智能时代T2Foundation Models

IT之家 · 智能时代 · 10/10/2026, 07:12:53

Microsoft Launches Decision-1: Post-trained on Qwen3.5-9B for Top Accuracy and Speed

By IT之家Original title: 微软推出 Microsoft-Decision-1 决策模型，基于 Qwen3.5-9B

78AI Score

Executive Summary

Microsoft has released Microsoft-Decision-1, a model specialized for structured decision-making tasks via post-training on Qwen3.5-9B. It achieved top accuracy in blind evaluations of nearly 150k samples and demonstrated response speeds 35x faster than GPT-6 Sol while maintaining high safety standards.

SOURCE COVERAGEFull Original Report

On October 10, it was reported that on October 9 (local time), Microsoft announced the launch of its decision model, Microsoft-Decision-1.

Microsoft launches Microsoft-Decision-1 decision model, based on Qwen3.5-9B

According to official statements, this model is specifically designed for structured decision-making tasks, delivering top-tier performance in both response latency and decision quality. Its actual performance comprehensively surpasses large language models (LLMs) and other traditional decision models. The model is now available on Microsoft Foundry and has been integrated into OpenRouter.

In a comparative evaluation involving 36 benchmarks and nearly 150,000 training-stage items under fully isolated conditions (Blind Evaluation), Microsoft-Decision-1 achieved the highest accuracy.

Additionally, benchmark data indicates that its response speed ranks first: it is 4.5 times faster than Quyet-1.0-Large, which ranked second, and 35 times faster than GPT-6 Sol.

Microsoft launches Microsoft-Decision-1 decision model, based on Qwen3.5-9B

During the development of Microsoft-Decision-1, Microsoft performed post-training on Qwen3.5-9B, achieving an ultra-low-latency single-inference decision scoring mechanism. This foundation will subsequently be adapted to other base models, including those from Microsoft AI (MAI) and OpenAI.

Furthermore, Microsoft conducted a systematic evaluation of Microsoft-Decision-1 based on 11 security benchmarks comprising a total of 5,250 requests. The testing scope covered harmful content, jailbreak attacks, and prompt injection. The results demonstrated that the model effectively intercepts various malicious behaviors while maintaining high usability.

Regarding pricing, input costs $0.042 per million tokens (Note: approximately 0.28 RMB at current exchange rates), while output is free.

Microsoft launches Microsoft-Decision-1 decision model, based on Qwen3.5-9B