微软发布了 Microsoft-Decision-1，这是一个小型模型，它在固定答案选项中进行选择，而不是编写文本。根据 Foundry 模型卡片显示，微软基于阿里巴巴的开源权重模型 Qwen3.5-9B 构建了该模型，随后对其进行了后训练。
Microsoft has released Microsoft-Decision-1, a small model that picks between fixed answer options instead of writing text. Microsoft built it on Alibaba’s open-weight Qwen3.5-9B and then post-trained it, according to the Foundry model card.
微软首席技术官办公室软件工程副总裁阿奇姆·斯里瓦斯塔瓦（Achint Srivastava）于 10 月 9 日在一篇博客文章中宣布了该模型。他写道，微软很快将基于其他模型对其进行重新构筑，包括 OpenAI 的模型以及微软自身的 MAI 模型。文章并未说明原因。
Achint Srivastava, a vice president of software engineering in Microsoft’s Office of the CTO, announced the model in a blog post on 9 October. Microsoft will soon rebase it on other models, including OpenAI’s and its own MAI models, he wrote. The post does not say why.
“决策模型正在迅速成为人工智能领域一个重要的新类别，”斯里瓦斯塔瓦写道。
“Decision models are quickly emerging as an important new category in AI,” Srivastava wrote.
给定一组固定的选项，Microsoft-Decision-1 会为每个选项返回一个校准的概率得分。它能够处理是或否的问题、多项选择以及评分。它还可以对照评分标准对人工智能的回复和智能体操作进行打分。
Given a fixed set of options, Microsoft-Decision-1 returns a calibrated probability score for each one. It handles yes or no questions, multiple choice and ratings. It can also grade AI responses and agent actions against a rubric.
模型卡片显示，微软列出了诸如路由、分类、优先级排序、验证和工作流控制等用途。该模型仅接受文本输入，并拥有 32,000 个标记的上下文窗口。卡片补充说，它不应成为关于人员的重大决定（例如信贷或就业）的唯一依据。
Microsoft lists uses such as routing, classification, prioritisation, verification and workflow control. The model takes text only and has a 32,000-token context window, the model card says. It should not be the sole basis for consequential decisions about people, such as credit or employment, the card adds.
微软表示，在包含 36 个基准测试的对比中，该模型的准确率最高。这涵盖了近 15 万个在训练中未公开的问题。
Microsoft’s own benchmarks Microsoft said the model scored the highest accuracy in its 36-benchmark comparison. That covered nearly 150,000 questions kept blind from training.
在其测试中，该模型的速度是排名第二的 H2O-Lightning-4B v1.1 的 2.5 倍。它的速度是 OpenAI 的 GPT-6 Sol 的 35 倍。当微软重新表述或重新排序相同的请求时，该模型平均有 1.3% 的时间会改变其决策。
In its tests, the model was 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1. It was 35 times faster than OpenAI’s GPT-6 Sol. When Microsoft reworded or reordered the same request, the model changed its decision 1.3% of the time on average.
微软还在 11 个安全基准测试的 5,250 个请求上对其进行了测试。这些测试涵盖有害内容、越狱尝试和提示注入。
Microsoft also tested it on 5,250 requests across 11 safety benchmarks. These covered harmful content, jailbreak attempts and prompt injection.
Xbox Research 使用该模型将 10,000 多条玩家反馈和评论按主题进行分类。微软表示，该模型在质量上与 GPT-6 Sol 具有竞争力，速度快了 14 倍以上，成本便宜了 200 倍。Copilot 团队发现它与 GPT5.6 Luna 具有竞争力，且速度快了 100 倍。
Internal tests and pricing Xbox Research used the model to sort more than 10,000 pieces of player feedback and reviews into themes. Microsoft said it was competitive with GPT-6 Sol on quality, over 14 times faster and 200 times cheaper. The Copilot team found it competitive with GPT5.6 Luna and 100 times faster.
输入成本为每百万标记 0.042 美元，输出标记免费。该模型可在 Microsoft Foundry 以及 Stripe 于 8 月收购的模型路由器 OpenRouter 上获取。
Input costs $0.042 per million tokens, and output tokens are free. The model is available in Microsoft Foundry and through OpenRouter, the model router Stripe bought in August.
微软随后更新了其博文，添加了来自 TypeSafe AI 的 Jev 的基准测试。正如托马斯·克拉布恩（Thomas Claburn）为《The》所报道的那样，Jev 在微软模型发布前三周推出。它成为了今年 Vercel 上采纳速度最快的模型。报道称，微软表示在其延迟测试中，其模型的速度是 Jev 的 2.8 倍。
Microsoft later updated its post to add benchmarks for Jev, from TypeSafe AI. Jev launched three weeks before Microsoft’s model, as Thomas Claburn reported for The. It became the fastest-adopted model on Vercel this year. Microsoft said its model was 2.8 times faster than Jev in its latency test, The reported.
据《The》报道，OpenAI 的 Decisions API 目前处于公开测试阶段。Cloudflare、Strands、Liquid AI、Perplexity 和 Snowflake 也已宣布推出决策模型。该媒体称，目前有 100 多个此类模型在争夺关注。
OpenAI’s Decisions API is in public beta. Cloudflare, Strands, Liquid AI, Perplexity and Snowflake have also announced decision models, according to The. More than 100 such models now compete for attention, the outlet said.