Microsoft has released Decision-1, a small AI model designed to make fast, structured choices rather than write text. It is built on Qwen3.5-9B and meant for applications that pick among predefined options. Qwen is an open-weight model family from Alibaba.
Microsoft says Decision-1 was the most accurate model in its tests. All performance figures in this story come from the company's own benchmarks.
What a decision model does
A decision model doesn't produce open-ended answers. It takes a fixed set of options and returns a calibrated probability for each one through a structured API call. A calibrated probability is a confidence score that is meant to match how often the model is actually right.
Developers can use those scores to route a support ticket or check an agent's proposed action. Applications can also send uncertain results on for further review. Microsoft lists uses such as agent controls, model routing, intent analysis, data labeling and content classification.
Satya Nadella promoted the launch on X on October 9. He wrote that it "outperforming both LLMs and other decision models in latency and quality."
What Microsoft's tests showed
Microsoft says the model reached the highest accuracy in a 36-benchmark evaluation covering nearly 150,000 questions withheld from training. The Decoder reports 83.5% accuracy and 85 ms latency, ahead of the startup Jev's model.
Reports differ on the speed advantage. The Decoder says Decision-1 is 2.5 times faster than the runner-up, H2O-Lightning-4B. Stocktwits reports it ran 4.5 times faster than its closest competitor and about 35 times faster than large language models such as GPT-6 Sol. The sources don't explain the gap.
One report says Microsoft's Copilot team used the system for quality checks on agent responses, running up to 100 times faster than earlier setups.
The Decoder notes that Cloudflare's open-source Clef models, which are also based on Qwen, weren't part of the comparison.
Availability and price
Microsoft's Foundry blog describes Decision-1 as in public preview. TestingCatalog, however, says the official catalog lists it as generally available. The Decoder says the model is available through Foundry and OpenRouter, a service for accessing many models. TestingCatalog says OpenRouter support is still due soon.