Microsoft Releases Decision-1: A Decision Scoring Model Built on Qwen3.5

Microsoft has recently unveiled a new model called Microsoft-Decision-1 — one that works very differently from the text-generation LLMs we're used to. Designed specifically for decision-making tasks such as routing, classification, verification, and agent control, this model is built through post-training on Alibaba's Qwen3.5-9B and is now available on Microsoft Foundry and OpenRouter.

What Is a Decision Model?

Traditional large language models are built around text generation — you give them a prompt, and they return a continuous natural language output. A Decision Model works in a fundamentally different way: rather than generating free-form text, it takes a fixed set of answer options and returns a calibrated probability score for each one.

In other words, when a system needs to make a judgment call — "which service should this request be routed to," "what category does this content belong to," "is this answer correct" — a decision model directly outputs a confidence probability for each candidate, rather than producing text that needs to be parsed downstream. This design makes outputs more controllable, more quantifiable, and far better suited for embedding into automated pipelines.

The Value of Probability Calibration

One of the key selling points of Microsoft-Decision-1 is that it returns calibrated probabilities. In real engineering contexts, if a model outputs 90% confidence, it should be correct roughly 90% of the time — that's what calibration means.

For use cases like routing and verification, calibrated probabilities are especially valuable. Systems can use them to set thresholds: automatically act when probability exceeds a certain cutoff, and fall back to human review or alternative logic when it doesn't. Compared to parsing the text output of a generative model, this kind of structured probability output dramatically reduces integration complexity and makes decision-making more transparent and auditable.

Background: Probability Calibration is a core concept from statistical learning. Uncalibrated models often suffer from overconfidence or excessive conservatism — for example, a model might output 95% confidence on a prediction that's only correct 70% of the time. Calibration aligns predicted probabilities with actual frequencies. Common methods include Platt Scaling and Temperature Scaling. Temperature Scaling is particularly simple and effective: by dividing logits by a temperature coefficient T before the softmax layer, the output probabilities can be "flattened" or "sharpened" to correct systematic biases. For engineering scenarios that require automated thresholds, good calibration means operations teams can estimate false-positive rates directly from historical data — allowing them to configure human-fallback triggers rationally rather than relying on vague empirical tuning.

Why Qwen3.5-9B as the Base Model?

Notably, Microsoft didn't train this model from scratch. Instead, they chose Alibaba's open-source Qwen3.5-9B as the foundation for post-training. At 9 billion parameters, this is a mid-sized model that strikes a balance between capability and inference cost — well-suited for high-frequency decision scenarios, since tasks like routing and classification often need to run at massive scale with high concurrency in production. Larger models would bring prohibitive latency and costs.

This choice also reflects the maturity of the open-source model ecosystem: a leading tech company building a commercial product directly on top of another company's open-source base is becoming standard practice in the industry.

Background: Post-training refers to further training stages applied on top of a pre-trained model, targeting specific tasks or capability goals. This typically includes Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF/RLAIF). Compared to pre-training from scratch, post-training requires several orders of magnitude less compute, while significantly reshaping model output behavior — for example, transforming a general-purpose language model into one that outputs only structured probability distributions. This is why Microsoft could launch a specialized product based on Qwen3.5-9B relatively quickly: the language understanding and reasoning capabilities built up during pre-training are preserved, while the output format and decision preferences are redirected through post-training, dramatically shortening the development cycle.

Use Cases and Practical Significance

Based on the official positioning, Microsoft-Decision-1 targets four core scenarios:

• Routing: Deciding which component should handle a request in multi-model, multi-tool systems;
• Classification: Structurally categorizing input content;
• Verification: Determining whether an output or answer meets requirements;
• Agent Control: Making next-step action decisions within agent workflows.

These happen to be exactly the points in current AI agent systems where "determinism" is most critical. As multi-agent collaboration, tool-calling, and model routing architectures grow increasingly complex, a lightweight decision layer dedicated to "making judgments" is becoming indispensable. At its core, Microsoft-Decision-1 provides a reliable decision hub for complex AI systems.

Background: In Multi-Agent system architectures, the complexity of "agent control" deserves special attention. A typical agent workflow may involve multiple specialized sub-agents (e.g., code execution, web search, database querying) along with tool calls and memory retrieval. The coordination layer must judge at every step: does the current state satisfy a termination condition? Which tool or sub-agent should be invoked next? Does an intermediate result need verification before proceeding? Relying on generative models to output natural language and then parsing it introduces additional error risks and makes system behavior hard to trace and reproduce. Specialized decision models, by outputting probability distributions over discrete options, naturally suit these "finite action space" control problems — conceptually closer to the role of a Policy Network in reinforcement learning.

Conclusion

Microsoft-Decision-1 represents one direction in the increasing specialization of large model applications: let generation models generate, and let decision models decide. By returning calibrated probabilities rather than free text, it provides more controllable, more easily integrated infrastructure for routing, classification, verification, and agent control in AI systems. Built through post-training on Qwen3.5-9B and available on both Foundry and OpenRouter, it means developers can relatively quickly evaluate it within their existing workflows. For teams building complex agent systems, this category of specialized decision models is well worth watching.