Microsoft releases Decision-1 model based on Alibaba’s Qwen architecture

Microsoft has released Microsoft-Decision-1, a specialised artificial intelligence model designed to choose between fixed answer options rather than generating open-ended text. The model, announced by Achint Srivastava, a vice president of software engineering in Microsoft’s Office of the CTO, was introduced in a blog post on 9 October. It represents a shift towards narrow, task-specific AI tools for enterprise workflows.

Root cause

Microsoft constructed the model using Alibaba’s open-weight Qwen3.5-9B as its base, according to the Foundry model card. The company then applied post-training to adapt the foundation for decision-making tasks. Srivastava noted that decision models are emerging as an important new category in artificial intelligence, distinct from general text-generation systems. Microsoft plans to rebase this model on other architectures, including OpenAI’s models and its own MAI models, in the future.

Attack path

The model is designed for deterministic outputs, returning a calibrated probability score for each option in a fixed set. It handles yes-or-no questions, multiple-choice scenarios, and rating scales. According to the model card, it can also grade AI responses and agent actions against a specific rubric. Microsoft lists typical uses as routing, classification, prioritisation, verification and workflow control. The system accepts text-only input with a 32,000-token context window.

Affected versions

The current release is Microsoft-Decision-1, based specifically on Qwen3.5-9B. Microsoft stated it achieved the highest accuracy in a 36-benchmark comparison covering nearly 150,000 questions kept blind from training. In latency tests, it was 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than OpenAI’s GPT-6 Sol. When Microsoft reworded or reordered the same request, the model changed its decision 1.3% of the time on average.

Mitigation

Microsoft tested the model on 5,250 requests across 11 safety benchmarks covering harmful content, jailbreak attempts and prompt injection. The model card advises that it should not be the sole basis for consequential decisions about people, such as credit or employment. Input costs $0.042 per million tokens, with output tokens free. The model is available via Microsoft Foundry and OpenRouter, the model router Stripe acquired in August.

Why it matters

Organisations adopting narrow AI models must distinguish them from general-purpose systems. Decision models offer speed and cost efficiency for specific tasks like classification but lack the versatility of large language models. Teams should verify that the model’s fixed options align with their workflow requirements. It is crucial to understand that these models do not generate novel text but select from predefined choices. Users must monitor for edge cases where prompt reordering might influence outcomes, ensuring robust validation layers are in place for critical operations.

What to do and how to stay safe: Microsoft

• Verify that the model’s fixed output options align with your specific classification or routing requirements before deployment.
• Monitor model outputs for consistency, particularly when input phrasing or order varies, to ensure decision stability.
• Ensure that automated decisions from the model are reviewed by humans if they impact individuals, such as in hiring or credit scenarios.
• Keep an eye on vendor updates, as Microsoft plans to rebase this model on other architectures in the future.

Step-by-step guide: Password Policies Examples That Reduce Password Fatigue

General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How does Microsoft-Decision-1 differ from standard large language models?

It does not generate free text but selects from a fixed set of options, returning probability scores for each choice.

Is the model available immediately?

Yes, it is available in Microsoft Foundry and through OpenRouter, with input costs of $0.042 per million tokens.

What performance benchmarks did the model achieve?

It was 35 times faster than OpenAI’s GPT-6 Sol and scored highest in Microsoft’s 36-benchmark accuracy comparison.