Microsoft announced this week the Microsoft-Decision-1, an AI model built specifically for making structured decisions rather than generating free-form text. Unlike the large language models (LLMs) that have dominated headlines for years, Decision-1 does not write paragraphs or reason through complex topics. It receives a fixed set of options and returns a calibrated probability for each one. The result is an entirely new category of AI designed for tasks where every call represents a micro-decision the software can execute immediately.
The model was built on top of Qwen3.5-9B, an open-weight architecture developed by Alibaba Cloud, the cloud computing arm of Chinese tech giant Alibaba. Microsoft did post-training on that base to specialize it in decision scoring, and the result is available in public preview through Microsoft Foundry and, soon, via OpenRouter. The announced price is $0.042 per million input tokens — and output tokens are free.
The proposal is simple, but ambitious. In modern AI applications, many of the most frequent steps do not require text generation: classifying an input, selecting an option, assigning a score, verifying a condition, prioritizing an item, or controlling a workflow. Microsoft calls these "decision models," and says they are becoming an essential component in compositional AI systems, where different stages use specialized models for different tasks.
Decision-1 was tested across 36 benchmarks covering nearly 150,000 questions, according to Microsoft. In the tests, it achieved 83.5% accuracy and 85 milliseconds latency — the company claims it is 2.5 times faster than H2O-Lightning-4B v1.1, the runner-up, and 35 times faster than GPT-6 Solution. Speed is the differentiator: in agentic systems that execute hundreds of decisions per session, every millisecond adds up.
But what makes this launch particularly interesting is not just the performance — it's the context. The model Microsoft chose as a base is Chinese. Qwen is developed by Alibaba Cloud, and the first version of Decision-1 rests entirely on it. The company promises that the next iteration will be rebased on its own MAI (Microsoft AI) models and OpenAI models, but did not explain why it started with Qwen or why it decided to keep the Chinese base in the initial version.
This move does not happen in isolation. The "decision models" category is growing at an impressive rate. According to The Register, there are now over 100 models of this type listed on the Hugging Face Decision Index. H2O.ai launched the Lightning-4B, built on Qwen3.5-4B. Strands launched the Decider 2B. Liquid AI unveiled the d1. Perplexity launched its own Decisions API. Snowflake announced its own proprietary model. Surogate released the Rune. And there is Jev, from TypeSafe AI, which was the first to arrive with force and became the fastest-adopted model on Vercel this year.
What differentiates Decision-1 from a conventional LLM is that it does not hallucinate in the same sense. Decision models do not invent facts — they choose among given options. This makes them suitable for critical tasks such as security incident routing, content classification, data validation, and judging AI agent responses. A language model might draft a response while Decision-1 evaluates whether that response meets defined criteria — a division of labor that is rapidly spreading across the industry.
Achint Srivastava, vice president of software engineering in Microsoft's Office of the CTO, emphasized in a blog post that "decision models are quickly emerging as an important new category in AI." He observed that with agentic AI becoming reality, cost plays a fundamental role in decisions about how to use AI models, and it is increasingly important to choose the right model for the right job.
The irony, however, is that the very logic driving decision models — choosing the right model for the right job — should, in theory, lead Microsoft to use its own MAI models from the start. Instead, the company started with Alibaba's Chinese Qwen and promises to migrate to its own technology afterward. The decision may be pragmatic: Qwen3.5-9B is competent, open-weight, and already available. But it reveals something about the current reality of AI: even OpenAI needs Chinese models to build its own tools, and China dominates a fundamental layer of the open-weight ecosystem upon which all major companies depend.
What follows will be interesting. Will Microsoft's promise to rebase on MAI and OpenAI models be fulfilled? And what does this mean for the decision models ecosystem, where competition is still forming? With more than 100 models in the market, the next year will likely bring consolidation — and the winners may not be the ones that look strongest today. The question is whether Microsoft, by starting with Qwen, has simply reinforced Alibaba's position as a foundational supplier for the global AI infrastructure — or if the future migration to MAI signals that the company is already planning independence.
Sources: Microsoft Foundry Blog, The Register, The Next Web
✓ Independent sources cross-checked and verified before publishing
