Microsoft releases a Qwen-based decision model that returns probabilities for fixed options

Take a team that routes support tickets with a large general model. Each ticket goes out, a paragraph comes back, and code parses it to find out whether the answer was billing or outage. Teams tend to tell themselves the router needs a model that understands the ticket. What the step does is pick one label.

Microsoft-Decision-1 picks between fixed answer options instead of writing text ^[4]. Given a fixed set of options, it returns a calibrated probability score for each one, and it can grade AI responses and agent actions against a rubric ^[6].

Achint Srivastava, the Microsoft vice president who announced the model ^[1], framed it as a category. "Decision models are quickly emerging as an important new category in AI," he wrote ^[3]. The model card is more specific. Microsoft lists routing, classification, prioritisation, verification and workflow control as uses ^[7]. The model takes text only and has a 32,000-token context window ^[8]. The card also warns that it should not be relied on by itself when a decision about a person carries real consequences, credit or employment for example ^[9].

The speed claims use a different yardstick each time. By Microsoft's count, it beat the second-placed H2O-Lightning-4B v1.1 on speed by a factor of 2.5 and OpenAI's GPT-6 Sol by a factor of 35 ^[11]. On Xbox Research's sorting of more than 10,000 pieces of player feedback, Microsoft said it was competitive with GPT-6 Sol on quality ^[13]. It was also over 14 times faster and 200 times cheaper, Microsoft said ^[13]. The Copilot team found it competitive with GPT5.6 Luna and 100 times faster ^[14]. Microsoft later updated its post to add Jev, from TypeSafe AI ^[18], and said its model was 2.8 times faster than Jev in a latency test, The Register reported ^[20]. Jev launched three weeks before Microsoft's model and became the fastest-adopted model on Vercel this year, according to Thomas Claburn's report in The Register ^[19].

For a router owner, the stability figure is 1.3%. That is how often, on average, the model changed its decision when Microsoft reworded or reordered the same request ^[12]. On 1,000 reworded tickets that is about 13 flipped decisions ^[22]. It is an average, so a given team's tickets could flip more or less often.

Calibrated probabilities are what a team sets thresholds on, for instance sending anything under 0.9 to a person. In our view a threshold tuned on the Qwen-based version should be rechecked once the planned rebase lands ^[2], since the rebase swaps the model that produces the scores.

To decide where the model belongs, sort each model call in your pipeline on two questions. The first is whether the answer is one of a list you can write down today. The second is what a wrong answer costs and whether anyone can reverse it. A fixed list with cheap, reversible errors, such as ticket routing or grading a draft agent action before a person sees it, is the cell to test. A fixed list with costly errors, anything touching credit or employment for a person, is where the model card says not to rely on the model alone ^[9]. An open-ended answer stays with a general model whatever the errors cost. The test for the first cell is a labeled sample of your own tickets, reworded, with the flip count compared against the 1.3% Microsoft reports ^[12].