最后更新:10/11/2026最后核验:2026-10-10

Microsoft-Decision-1

Microsoft-Decision-1

4
0

微软的小型决策评分模型,已在 Microsoft Foundry 和 OpenRouter 提供:对文本或 JSON 状态回答是/否、选择、评分等类型化问题,返回校准概率而非生成文本,适用于分类、路由、排序、核验与工作流控制。

付费
广告

<p>Microsoft-Decision-1 面向需要确定性、可校准决策输出的应用。与生成式大模型不同,它不生成自由文本或理由,而是读取提供的 state(文本或 JSON)和一组带类型的 questions,为每个固定选项返回 0 到 1 的校准概率,应用可据此决定执行、搁置还是转人工。问题类型包括 noul(是/否,用于门控与过滤,如工具调用是否触及生产数据、回复是否承诺退款)、choice(从固定选项集中选一,用于工单分类与模型/工具选择)和 score(有序刻度评分,用于优先级排序)。它已在 Microsoft Foundry 上线,支持 Microsoft Entra ID 或 API key 认证,并可通过 OpenRouter 的 Decisions API 调用(输入 $0.042/百万 token、输出免费)。该模型基于阿里开源 Qwen3.5-9B 后训练而成,权重持续更新但 API 形状保持稳定。注意:标准 chat completions SDK 不适用,概率阈值应在标注样本上验证后再用于自动化门控。</p><ul><li>为固定选项返回校准概率,而非生成自由文本</li><li>noul/choice/score 三类问题覆盖门控、分类、路由与评分</li><li>Microsoft Foundry 部署使用,OpenRouter 输入 $0.042/1M、输出免费</li><li>基于 Qwen3.5-9B 后训练;面向分类、路由、核验、AI judging 与护栏</li></ul>

详情

平台

API

功能特性

  • Returns calibrated probability per fixed answer option instead of free-form generated text
  • Typed noul (yes/no), choice, and score question types for gates, filters, routing, and grading
  • Available in Microsoft Foundry with Entra ID or API-key auth, and through the OpenRouter Decisions API
  • Post-trained from Alibaba open-weight Qwen3.5-9B for single-pass scoring; weights update while the API shape stays stable

支持语言

en

适合人群

Applications that need predictable, typed decision outputs with confidence thresholdsSupport-ticket, incident, and content triage flows: classify, route, prioritize, and filterAgent guardrails and AI judging: verification, prompt-injection checks, and model/agent selection

已知限制

  • Not intended for open-ended generation, conversation, translation, or summarization
  • Uses the OpenRouter Decisions API rather than the OpenAI-compatible chat endpoint, so standard chat SDKs do not work
  • OpenRouter pricing is $0.042 per 1M input tokens with output free; throughput is capped by the single Azure provider
  • Probability thresholds for automated gates should be validated against labeled examples before use

替代工具

Jev — TypeSafe decision API that also returns typed probability-calibrated answers
OpenAI Structured Outputs — General LLM API for structured JSON when decisions also need reasoning text
Anthropic Claude API — General model API for richer classification workflows with tool use

为此工具评分

相关工具