Microsoft AI Releases Microsoft-Decision-1: A Qwen3.5-9B Decision-Scoring Model

Microsoft has a new model that doesn’t try to sound smart. It tries to choose. Microsoft-Decision-1 is built for routing, classification, verification and agent control, and it returns calibrated probabilities for fixed options instead of generating prose. The company says that’s the point: software can act on a score much more cleanly than on a paragraph.

The model is post-trained from Alibaba’s Qwen3.5-9B and is available now in Microsoft Foundry and through OpenRouter. Microsoft says future versions will be based on its own MAI models and OpenAI models. The supported use cases in the Foundry card are narrow and practical: yes/no calls, multiple choice, ratings, classification, rubric-style grading of AI responses, groundedness checks against evidence, and explicit abstentions like “cannot tell.” Output is JSON. No explanations. No text generation.

On Microsoft’s own benchmark run, Decision-1 came out on top on average accuracy across 36 benchmarks, reaching 83.5%. The company says it tested 9 systems on 147,137 questions, with the benchmark set kept blind from training. It also claims a p50 latency of 85 ms and p95 of 125 ms. Microsoft says that’s 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol, which it says took 3.01 seconds.

The catch is that this is a closed API model, not open weights. Microsoft lists no explanations, no images, no audio, no video, and no use for open-ended chat, translation or summarization. It also says the model should not be the sole automated decider for credit, employment, housing, healthcare or legal rights. So yes, this is a very specific product. That specificity is the whole pitch.