Microsoft决定1:9B Qwen型号发布解释

中文版正在随流水线自动整理，稍后刷新即可；可先阅读下方英文原文。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Microsoft used October 9, 2026 to introduce a model that refuses to write you a paragraph. Microsoft-Decision-1 doesn’t chat, summarize, or draft emails. It scores a fixed list of answer options in a single forward pass and hands back a probability for each one, then gets out of the way. The company built it on top of Alibaba’s Qwen3.5-9B, made it available through Microsoft Foundry immediately and through OpenRouter shortly after, and is already pointing to internal uses ranging from incident response to scientific discovery.

It’s a narrow launch by design, and that’s exactly what makes it interesting. While the rest of the industry races to ship bigger, chattier assistants, Microsoft spent its October announcement on a model built to do one thing: pick an answer from a list, fast, and tell you how confident it is. For engineering teams building agent pipelines, routing layers, and automated verification systems, that’s arguably more useful than another general-purpose chatbot.

Don't miss new tech stories on Google

Add Shattered once in the Google app and our stories show up in your news suggestions.

What Microsoft-Decision-1 Actually Does

Most AI headlines this year have been about models that generate more: more tokens, more context, more creative output. Microsoft-Decision-1 moves in the opposite direction. According to the company’s own announcement, the model 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. That’s a meaningfully different job than what GPT-6 Sol, Gemini 4 Argon, or Claude Opus 5.5 are built to do.

Microsoft named four core use cases for the model at launch: routing, classification, prioritization, and verification, plus what it calls workflow control. Think of a support ticket that needs to be routed to the right team, a piece of user content that needs a yes/no moderation call, or an AI agent’s output that needs a pass/fail check before it’s allowed to proceed. Those are high-volume, low-latency decisions, and they’re exactly the kind of task that a 9-billion-parameter specialist model can do more cheaply than a full-sized reasoning model.

The distinction matters for anyone building with large language models today. A chatbot generates a response and you have to parse it, guess at its confidence, and hope it followed your formatting instructions. A decision-scoring model skips that parsing step entirely: the output format is already a probability distribution over options you defined. That’s a smaller, more mechanical problem, and it’s one Microsoft is betting plenty of production systems have in volume.

Inside the Model: Why Qwen3.5-9B as the Base

The detail that generated the most commentary wasn’t the use case, it was the foundation. Microsoft-Decision-1 is post-trained from Qwen3.5-9B, a 9-billion-parameter open-weight model from Alibaba. The Register flagged the choice directly, framing the launch as Microsoft leaning on a Chinese AI lab’s open-weight release to get the product out the door quickly, while noting Microsoft’s own stated intent to migrate the model’s foundation over time.

Microsoft has been explicit that this is a starting point, not a destination. In its launch post, the company said: “To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.” That’s a notable admission for a company that has poured billions into its own model development through the Microsoft Foundry platform and its OpenAI partnership. It suggests the Qwen base was chosen for speed to market on a narrow, well-defined task, not as a long-term architectural commitment.

Using someone else’s open-weight model as scaffolding isn’t unusual in 2026. Plenty of smaller, task-specific tools get built this way, the same logic that drove projects like JetBrains’ Mellum2.1 coding model to pick a compact parameter count over raw scale. What’s notable here is that it’s Microsoft, a company with its own frontier model ambitions, publicly choosing a rival lab’s weights for a flagship-named product, then telling customers up front that the underlying engine will change later.

The Benchmark Claims: 36 Tests, Nearly 150,000 Questions

Microsoft is leaning hard on numbers to justify the product. The company says Decision-1 achieved the highest accuracy across a comparison spanning 36 benchmarks and approximately 150,000 held-out questions, a scale meant to signal breadth rather than a single cherry-picked test. Microsoft reported 83.5 percent accuracy across that evaluation, a figure the German outlet BornCity also relayed.

Speed claims followed the same pattern of big, specific numbers. Industry newsletter AI Weekly and the AI-tracking site TestingCatalog both reported that Microsoft claims Decision-1 runs 4.5 times faster than a model it calls Quyet-1.0-Large, and 35 times faster than GPT-6 Sol. A separate write-up from Data Studios led with the 35x figure as its headline framing. Microsoft’s own launch material describes the GPT-6 Sol result as approximately 35× lower P50 latency, measured on its own benchmark, and says Decision-1 was 2.5× faster than H2O-Lightning-4B v1.1, the runner-up in its comparison. That is a specific P50 measurement rather than a general-purpose speed advantage, since the systems perform different tasks: Decision-1 scores fixed options while GPT-6 Sol generates text. Taiwan’s Liberty Times reported response times in the neighborhood of 85 milliseconds, though it didn’t specify the hardware or test conditions behind that number.

Here’s where a healthy dose of skepticism belongs. These are Microsoft’s own benchmark results, reported through the company’s launch materials and relayed by outlets covering the announcement. None of the coverage reviewed for this piece includes an independent, third-party reproduction of the 36-benchmark comparison, the 4.5x and 35x speed multipliers, or the 85-millisecond latency figure. That doesn’t make the numbers false, but it does mean they should be read as vendor-reported performance claims rather than peer-reviewed results, at least until someone outside Microsoft runs the same tests.

One figure got less attention but may matter more in production: Microsoft says the model’s 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 a real concern for any model making binary or multi-option calls that downstream systems act on automatically. If a routing decision flips because of a trivial rewording, that’s a production bug waiting to happen, so a high stability number, if it holds up under outside testing, would be a genuinely useful data point for engineering teams evaluating the model.

Pricing: $0.042 per Million Input Tokens, Free Output

Microsoft’s own announcement and Microsoft Foundry documentation don’t publish a confirmed price in the sources reviewed for this article. What’s circulating instead comes from OpenRouter’s own model listing and was picked up by AI Weekly and the Italian outlet Pasquale Pillitteri: $0.042 per million input tokens, with output free. That pricing structure makes sense for a model whose entire output is a short list of probability scores rather than generated text, since there’s effectively nothing to meter on the output side.

At that price, Decision-1 is positioned as a genuinely cheap option for high-volume routing and classification work, the kind of task that currently gets routed to smaller general-purpose models mostly because nobody wants to pay full-model prices to ask “is this spam, yes or no.” It’s a similar cost logic to what pushed Anthropic to cut Claude Haiku pricing by 75 percent earlier this year: cheap, fast models win the high-frequency, low-complexity slice of the API traffic pie, even when a flagship model would technically do the job too.

Availability: Foundry First, OpenRouter Close Behind

Microsoft-Decision-1 launched in public preview on Microsoft Foundry, where documentation describes it as available for classification, routing, ranking, grading, and binary decisions. Microsoft’s launch post initially described OpenRouter availability as “coming soon,” but OpenRouter’s own model listing confirmed the model had gone live under the identifier microsoft/microsoft-decision-1 within roughly a day of the announcement, with OpenRouter posting that it was “live on OpenRouter.”

Weights for the model aren’t being distributed, based on reporting from Data Studios, meaning developers access Decision-1 exclusively through hosted APIs rather than downloading and self-hosting it, a contrast with the fully open Qwen3.5-9B it’s built on. That’s a common pattern for commercial derivatives of open-weight models, but it’s worth flagging for teams that assumed a Qwen-based product would come with the same openness as the base model.

Working With Decision-1 in Foundry

The model accepts text and supports inputs up to 32,000 tokens. It returns calibrated scores for predefined options, and it does not generate explanations or handle images, audio, or video. Teams that need a written rationale or multimodal input will have to pair it with a general-purpose model rather than expect Decision-1 to cover that work.

The confidence score is the part of the output that applications act on. The probability itself is part of the API, not just a ranking score, and applications use confidence to decide when to act, defer, or ask for review. One routing example reads: if urgent is above 0.9 and priority is above 1.5, page the on-call engineer immediately. Another example sets a floor: if the top probability is below 0.65, a local policy can avoid downgrading, keep the current model or request manual selection.

A worked example in the documentation shows the model returning a weighted priority score of 1.9868 with 98.08% confidence. A 90% prediction should be right about nine times out of 10 on representative cases, which is the standard teams should hold the scores to when they test them on their own data. The quickstart needs Python 3.10 or later, so a first pilot can be small. Because Microsoft describes pricing as dynamic, it is worth checking the current Foundry listing before sizing a deployment.

Nadella’s Pitch and Microsoft’s Internal Use Cases

Satya Nadella personally introduced the model on X, writing: “Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.” That’s a specific, named set of internal applications, three in total, and it signals that Microsoft isn’t just shipping Decision-1 to third-party developers. The company is running it through its own operational workflows first.

Incident response is a logical first stop: a decision-scoring model could triage alerts, score severity, or decide whether a ticket needs to escalate to a human, all tasks that fit the “fixed options plus confidence score” shape perfectly. Quality control and scientific discovery are broader claims, and no named Microsoft executive beyond Nadella has gone on record describing how those workflows actually use the model, although Microsoft-published material has since described other internal trials in more detail.

Early Internal Trials and Example Decisions

Microsoft reported that Xbox Research used Decision-1 to categorize more than 10,000 pieces of open-ended feedback into predefined themes, with quality described as competitive with GPT-5 and execution reported as 80–100 times faster. Other Microsoft-published material says teams also evaluated the model for Copilot-response assessment and adaptive r

Auto-fetched and translated by AI Atlas. All rights belong to the original author.