Microsoft launches Decision-1, a 9B routing model built on Alibaba's Qwen3.5

Satya Nadella introduced Decision-1 as a model for 'fast decision-making.' He claims it beats both LLMs and other decision models on latency and quality for structured choices. Microsoft is testing it internally for incident response, quality control and scientific discovery.

The design is the interesting part. Decision-1 does not generate free text. It returns a calibrated probability distribution over options the developer defines in advance. That makes it a natural fit for agent routers (which tool or model to call next), triage queues and verification steps, where you need a confidence score rather than a sentence.

The Qwen base is notable. Microsoft chose an Alibaba open-weight model as the foundation for a first-party product, underscoring Qwen's position as a default open base even for US hyperscalers. Alongside Claude in Copilot and in-house Maia silicon, developers see it as more evidence of Microsoft diversifying away from OpenAI.

The main complaint from builders is that the weights are hosted-only despite the open Qwen base, which rules out self-hosting for latency-sensitive or air-gapped routing. Hugging Face has meanwhile added hub tags for decision models, a sign the category may be forming.