Microsoft releases Decision-1: a Qwen3.5-9B-based model that picks an answer from a list

Microsoft has released Microsoft-Decision-1, a compact 9-billion-parameter model. It does not generate text. Instead, it picks one answer from a predefined set. The base is Alibaba's open Qwen3.5-9B, which Microsoft fine-tuned in a post-training stage. This is according to the model card in Foundry. The release was announced in a blog post by Achint Srivastava, the company's vice president of software engineering in the Office of the CTO.

What exactly Microsoft built

Decision-1 solves a narrow task. The input is a situation and a list of allowed options; the output is one of them. There is no free text: the answer format is tightly constrained.

Models like this are useful where ordinary software code is built around the AI. Typical uses include routing requests, choosing an agent's next step, classification, and ranking options. A large generative model can handle these tasks, but it costs more and responds less predictably. A 9-billion-parameter model with a constrained output is cheaper to run and easier to verify: the answer is either in the list or it isn't.

This is our own conclusion about typical use cases for this class of model. For Decision-1's specific applications, check Microsoft's announcement itself.

Why Qwen instead of Microsoft's own Phi?

Microsoft has its own family of small models, Phi. Even so, for Decision-1 the company took Alibaba's open weights. That says something about how the small-model market works: teams increasingly take a ready-made base model, and the value is created in fine-tuning for a specific task. For a large American company, using a Chinese open-weight model in a product on its own platform is also a signal. Here, the quality and openness of the base outweighed its origin.

There is no sign that Microsoft tried to hide the base model. The source is stated directly in the model card.

Key specs

• Name: Microsoft-Decision-1.
• Base model: Alibaba's Qwen3.5-9B, open-weight.
• Size: about 9 billion parameters.
• Output type: a choice from fixed options, not text generation.
• Tuning: post-training by Microsoft.
• Platform: the Foundry model catalog.
• Announcement: Achint Srivastava, Office of the CTO.

The available portion of the announcement contains no benchmarks, no description of the training data, and no license terms. Without them, there is nothing yet to compare Decision-1 against general-purpose models.

What to check before adopting it

If your product has a "pick one of N" step, Decision-1 is worth testing alongside your current solution. Compare four things: accuracy on your own data, latency, cost per request, and consistency of answers across repeated runs. Expect an advantage specifically where the options are known in advance and there are few of them.

Separately, read the model card in Foundry. You need to confirm the license, which is inherited from Qwen3.5 and may carry its own terms. It is also worth finding out whether the model supports your language and the format of your option list. If the answers to those questions work for you, a small specialized model can replace a large LLM call in part of your pipeline and cut costs without losing controllability.