Microsoft Decision-1: 9B Qwen Model Launch Explained

Microsoft used its October 9, 2026 announcement to launch Decision-1, a small AI model that does not write open-ended prose. According to the company, it takes content plus a fixed set of answer options and returns calibrated probability scores for each option in one forward pass. The launch matters because it signals a shift toward narrow, decision-focused AI rather than another general-purpose chatbot.

What Microsoft-Decision-1 Actually Does

Microsoft Decision-1 does not chat, summarize, or draft. According to the source, it reads supplied content alongside a set of predefined answer options and returns a calibrated probability score for each one. Instead of generating open-ended text, it produces a probability distribution over options a developer defines. That output format is meant to remove parsing work from downstream systems.

The company named routing, classification, prioritization, verification, and workflow control as core use cases. Examples include support ticket routing, yes or no moderation, and checking an AI agent output before it proceeds. For engineering teams, the appeal is high-volume, low-latency decisions. A chatbot generates a response, and a developer must parse it, guess at confidence, and check formatting. A decision model returns probabilities directly, which is a smaller mechanical problem.

The distinction is important for anyone building with large language models. A general assistant may be creative, conversational, or useful for drafting, but Decision-1 is specialist. It is built for tasks where the answer space is known and the system needs a fast confidence signal. That makes it relevant to routing layers, agent pipelines, and automated verification systems. The source also notes the model does not handle images, audio, or video, so teams needing multimodal input or written rationale would need another model.

Why Microsoft Built It on Qwen3.5-9B

The most discussed technical detail was the base model. Microsoft said Decision-1 was post-trained from Alibaba Qwen3.5-9B, a 9-billion-parameter open-weight model. The company said it post trained Qwen3.5-9B for fast, single-pass decision scoring and plans to rebase future versions on other models, including Microsoft AI and OpenAI.

Microsoft described the Qwen base as a starting point, not a final architecture. Reporting on the launch noted that Microsoft chose a Chinese AI lab open-weight release to move quickly on a narrow task. The source says Microsoft was explicit that the foundation would migrate over time.

For builders, the choice raises practical questions. A post-trained open model can be adapted for a specific output format, but it may not match a fully Microsoft-built model. The company's plan to rebase suggests Decision-1 is a research and product experiment, not a permanent bet. That is a useful frame for teams evaluating whether to adopt the system now or wait for the next foundation. The launch therefore became as much about product strategy as architecture.

What Microsoft Claimed on Benchmarks and Pricing

Microsoft claimed Decision-1 performed best across 36 benchmarks and about 150,000 held-out questions, reporting 83.5 percent accuracy. The company also claimed it ran 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol, describing the latter as approximately 35 times lower P50 latency. It said Decision-1 was 2.5 times faster than H2O-Lightning-4B v1.1. A report mentioned response times near 85 milliseconds.

These figures come from Microsoft materials and coverage of the announcement. The source notes that no independent third-party reproduction was reviewed. The claims are vendor-reported performance numbers, not peer-reviewed results, at least until outside testing occurs. The accuracy figure came from the company's evaluation, and the speed numbers were measured by Microsoft. The comparison included models performing different tasks, which complicates direct claims.

One figure that may matter for production is perturbation stability. Microsoft said decisions were perturbation-stable 98.7 percent of the time, according to the company's launch summary cited by AI Weekly. Stability under small input changes is important when systems act automatically on probabilities. If a routing decision flips because of a minor change, that can become a production issue.

On cost, Microsoft did not publish a confirmed price in the reviewed material. OpenRouter's model listing put the price at 0.042 per million input tokens, with free output. The source said that pricing structure fits a model whose output is mostly probability scores rather than generated text.

Availability and How to Work With It

Decision-1 launched in public preview on Microsoft Foundry, where documentation described it for classification, routing, ranking, grading, and binary decisions. OpenRouter said the model was live under microsoft/microsoft-decision-1 within about a day. Weights were not described as distributed, so access appeared to be through hosted APIs rather than self-hosting, unlike the open Qwen base.

The model accepts text and supports inputs up to 32,000 tokens. It returns calibrated scores for predefined options and does not generate explanations or handle images, audio, or video. The confidence score is part of the API output, and applications can use it to act, defer, or ask for review. A routing example in the source described thresholds, such as requiring urgent above 0.9 and priority above 1.5 before paging an on-call engineer.

A practical workflow would pair Decision-1 with a general model for tasks that need explanation or multimodal input. The source notes that the decision model does not cover those areas. Teams evaluating it should test the confidence thresholds against their own data and latency requirements.

How This Relates to Grafy

For creators, developers, and researchers, the story points to a broader trend: AI systems may split into general assistants and specialized decision engines. Grafy's Grafy Chat is a multi-model chat surface, but the source does not state that Decision-1 is available there. The genuine connection is the work of comparing models, routing queries, and choosing the right model for the task.

Readers interested in the cost side of model selection can also look at lowest-cost model routing. The Decision-1 launch does not prove that a small decision model beats all general models for every use case, but it adds another data point in the push toward cheaper, faster, task-specific AI.

Frequently asked questions

What is Microsoft Decision-1?

Microsoft Decision-1 is a small AI model that does not generate open-ended text. According to the source, it takes content and predefined options and returns calibrated probability scores. It was built for routing, classification, prioritization, verification, and workflow control. The launch focused on fast, high-volume decisions rather than chat.

How fast and how much does Decision-1 cost?

Microsoft claimed Decision-1 was about 35 times faster than GPT-6 Sol and 2.5 times faster than H2O-Lightning-4B v1.1 in its own tests. A reported cost was 0.042 per million input tokens with free output. The source noted Microsoft's price came from OpenRouter's listing, while speed claims need independent checks.

How could a developer try decision scoring with Grafy?

A developer could use Grafy Chat to compare small and large models on a routing or classification task, though Grafy may not currently list Decision-1. The idea is to test confidence thresholds and output formats. For related reading, see Grafy's lowest-cost model routing guide.

---

Source: Hacker News — front page — Microsoft Decision-1: 9B Qwen Model Launch Explained