Microsoft Unveils Decision-Centric AI Model
Microsoft has introduced Microsoft-Decision-1, an innovative artificial intelligence model designed to enhance automated decision-making processes. This new system specializes in tasks such as routing, classification, verification, and agent control by offering calibrated probabilities for a set of predefined answer options, rather than generating free-form text. The model is built upon Alibaba's Qwen3.5-9B architecture and is now accessible through Microsoft Foundry and OpenRouter.
Understanding Decision Models in AI
Decision models represent a distinct category within artificial intelligence, specifically engineered to process inputs and assign scores to a finite set of choices. Unlike large language models that produce explanatory prose or creative content, these models deliver outputs that software systems can immediately act upon. Microsoft-Decision-1 embodies this approach by focusing on providing a probability for each option in a single call, enabling direct integration into automated workflows.
Technical Specifications and Availability
- Based on the Qwen3.5-9B model, though specific parameter counts remain undisclosed.
- Features an expansive 32,768-token context window.
- Available exclusively as a hosted API through Microsoft Foundry and OpenRouter via Azure.
- There are no plans for open weights releases, quantized variants, or disclosure of specific underlying hardware.
Future iterations of the model are expected to incorporate foundational elements from Microsoft's own MAI and OpenAI models.
Performance and Internal Applications
Microsoft conducted extensive internal evaluations, comparing Microsoft-Decision-1 against nine other systems across 36 benchmarks comprising over 147,000 questions. The new model achieved the highest average accuracy at 83.5%, surpassing competitors like Quyet-1.0-Large, which scored 81.9%. It also demonstrated exceptional speed, with a median latency of 85 milliseconds. This performance is notably faster than Quyet-1.0-Large and significantly quicker than models such as GPT-6 Sol.
Internal use cases at Microsoft highlight the model's efficiency:
- Xbox Research: Categorized over 10,000 feedback items, operating 14 times faster and 200 times more affordably than GPT-6 Sol.
- Copilot Quality Control: Proved competitive with GPT5.6 Luna while delivering results 100 times faster.
- Microsoft Discovery: Exhibited 46 times greater consistency compared to traditional LLM scoring, at triple the speed.
Deployment, Pricing, and Limitations
Developers can integrate Microsoft-Decision-1 via Microsoft Foundry as a 'Direct from Azure' model. On OpenRouter, it is accessed through the Decisions API, distinct from chat endpoints. The pricing structure is competitive, with input tokens costing $0.042 per million, and output tokens being free. This contrasts with OpenAI's Luna decisions endpoint, which charges $0.10 per million input tokens.
It is important to note the specific limitations of Microsoft-Decision-1:
- It is not designed for text generation, open-ended questions, chat functions, translation, or summarization.
- The model processes text inputs only; it does not support images, audio, or video.
- It does not provide explanations or rationales for its decisions.
- The model is not intended for sole automated decisions in critical areas such as credit, employment, housing, healthcare, or legal rights.
- Applications employing the model must incorporate defined options, thresholds, escalation protocols, and human oversight.
While Microsoft's performance claims are robust, particularly regarding latency, some industry observers, including H2O.ai, have raised questions about the methodology of competitor latency comparisons. Independent benchmarking on platforms like JevBench is still pending for Microsoft-Decision-1.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
