Microsoft makes the fast Microsoft-Decision-1 model available

Microsoft made Microsoft-Decision-1 available on Microsoft Foundry. The model gives a score for each alternative in one step. Microsoft writes that it had the top accuracy in a test of 36 benchmarks with almost 150,000 questions. It also writes that the model is 4.5 times as fast as Quyet-1.0-Large and 35 times as fast as GPT-6 Sol.

What it does

The model is Qwen3.5-9B with more training by Microsoft. It gives a probability score for each alternative in a set. It can also give a score to AI responses and agent steps with a rubric.

Microsoft writes that the model is for these tasks:

• routing
• classification
• prioritization
• verification
• workflow control

How stable it is

Microsoft made 8 changes to the same input and counted the decisions that changed. The decision changed in 1.3% of the changes, on average. It did not change when Microsoft reversed or shuffled the alternatives.

Internal tests

Xbox Research used the model to put game feedback in groups of themes. Microsoft writes that the model was almost as good as GPT-6 Sol in quality. It was more than 14 times as fast, and the cost was 200 times less.

What comes next

After this, the model will be available through OpenRouter. Microsoft will also change the base of the model to other models, for example Microsoft AI models and OpenAI models.

This is a brief. We point to the report and do not rewrite it. Read it at the source below.