Microsoft Announces Qwen-Based Decision-Focused AI Model Microsoft-Decision-1

Summary

Microsoft announced 'Microsoft-Decision-1'. This fine-tuned model, based on Alibaba's Qwen3.5-9B with 9 billion parameters, returns calibrated probability scores for predefined options rather than generating free-form text. It charges $0.042 per million tokens for input, while output is free.

Why it matters

The model targets a limited but high-volume enterprise need: classification, routing, and monitoring agent actions. In these tasks, a full-fledged generative model unnecessarily consumes resources, making latency and cost critical. Microsoft offers the same price as TypeSafe AI's 'Jev' model while claiming to be 35 times faster than GPT-6 Sol. The company aims to secure a position in the rapidly evolving 'decision API' category, where OpenAI is also entering with its Decisions API preview by the end of September.

Highlights

• Microsoft fine-tuning on an open-weight 9B base (Qwen3.5-9B); context limit of 32,768 tokens, text input and output only
• The Xbox research team reported that it was 14 times faster than GPT-6 Sol and cost 1/200th as much in a feedback classification task with over 10,000 items
• Microsoft explicitly warns against using the model alone for credit, hiring, or medical decisions

Critical take

The highlighted speed and cost advantages were measured against GPT-6 Sol, a much larger and general-purpose model, making the comparison akin to apples and oranges. Moreover, in calibration accuracy—which could be considered the model's main selling point—Decision-1 ranked only third, behind Jev and other models. Since the entire architecture is built on a fine-tuned 9B model, the 'decision-specific' framework likely serves more as a marketing tool than a reflection of the underlying true capability.