最新模型
Microsoft Releases Microsoft-Decision-1, Scoring Option Probabilities for Agent Workflows in a Single Inference
Microsoft launched Microsoft-Decision-1 on October 9. Post-trained from Qwen3.5-9B, it returns probabilities for predefined options in classification, routing, and agent action evaluation. The model is listed on Foundry and OpenRouter, but its performance and confidence calibration still need validation against real workloads.
On October 9, Microsoft released Microsoft-Decision-1, a decision-scoring model designed for classification, routing, and action evaluation—tasks common in agent systems. Post-trained from Qwen3.5-9B, it assigns probabilities to fixed answer options in a single inference, supporting yes/no questions, multiple choice, scoring, and rubric-based assessments of answer quality. Official announcement
This interface lets applications read each option’s score directly and then decide, based on thresholds, whether to proceed, defer, or send the case for human review. For example, after an agent proposes its next action, the system could offer candidates such as “continue,” “retry,” “stop,” and “hand off.” The technical benefit is less text generation and parsing for workflow control. However, the application remains responsible for ensuring that the candidate actions are complete and the decision criteria are clear. This is an engineering implication inferred from the model interface, not a guarantee of overall agent reliability.
Microsoft says the model achieved the highest accuracy across 36 benchmark comparisons covering nearly 150,000 questions, with P50 latency about 35 times faster than GPT-6 Sol. Across tests with eight types of input perturbations, the average decision-flip rate was 1.3%. These are vendor-reported evaluations and cannot be directly generalized to Chinese-language data, different numbers of options, or real-world network conditions. Evaluation details
For deployment, the Foundry launch post lists the service as public preview. At the time of checking, OpenRouter listed Azure as a provider. Its model page specifies text input, a 32,768-token context window, a price of $0.042 per million input tokens, and free output tokens. It also notes that the weights will be updated over time while the API format remains consistent. Foundry launch post, OpenRouter model page
Engineering teams should next measure Traditional Chinese classification accuracy, consistency when options are reordered, and the rates of incorrect decisions and human handoffs at different confidence thresholds. Ongoing weight updates also mean behavior may change even when the model name stays the same, so teams should retain a fixed regression dataset. For tool permissions or safety decisions, probability scores can help route cases, but the system should still enforce authorization rules deterministically.