Microsoft Decision-1: 9B Qwen Model Launch Explained

Who Should Read

Backend developers who use expensive general LLMs for ticket routing or agent pass/fail checks and worry about cost and latency

Core Mechanics

• It scores options instead of writing text
  • Microsoft-Decision-1 takes content plus predefined options and returns a calibrated probability per option in a single forward pass.
  • Core uses are routing, classification, prioritization, verification and workflow control.
• The base is Qwen3.5-9B
  • Microsoft post-trained Alibaba's open-weight model for fast single-pass decision scoring.
  • It says it will later rebase on MAI and OpenAI models.
• Speed and accuracy are vendor benchmarks
  • Microsoft claims 83.5% accuracy across 36 benchmarks and about 150,000 held-out questions.
  • It claims roughly 35x lower P50 latency than GPT-6 Sol, though the tasks differ.
  • The article cites no independent reproduction.
• A stability number is reported
  • Microsoft says decisions were perturbation-stable 98.7% of the time.
  • This matters because routing that flips on trivial rewording is a bug, but it is unverified externally.
• Limits are clear
  • Input is text only, up to 32,000 tokens.
  • It generates no explanations and handles no images, audio or video.
  • Weights are not distributed, so access is via hosted API only.
• Pricing comes from OpenRouter's listing
  • Reported price is $0.042 per million input tokens with free output.
  • Microsoft's own docs show no confirmed price, and Foundry pricing is described as dynamic.

Community Reactions

• Opinion that it reads like an ad
  • One commenter said the article reads like promotion because it never mentions other decision models.
  • Another commenter pointed to Microsoft's own post (commandline.microsoft.com) as the real news.
• Experience with other open decision models
  • One commenter said pplx-decider-v1.2 and its -flash variant were pretty good in what they evaluated.
  • They evaluated on MonitoringBench and asked if anyone had compared these models on their own tasks.

How to Apply

• Test thresholds on 100 past tickets
  • Today, pick 100 tickets that humans routed and compare your current LLM's choices to the human decisions.
  • Set a confidence threshold such as 0.75 and measure agreement when auto-handling only above it.
• Split classification calls from your general LLM
  • List calls that need no generation, such as spam checks, A/B/C classification and agent pass/fail checks.
  • Point them at microsoft/microsoft-decision-1 on OpenRouter and compare cost and latency.
• Send low confidence to humans
  • If the top probability is below the threshold, route to a manual review queue instead of acting automatically.
  • Pair it with a general model when you need a written rationale.

Try It

# Paste into Claude or ChatGPT (simplified version of the Decision-1 idea: per-option probabilities only)

You are a classifier. Output JSON only, no prose.
Options: ["billing_support", "technical_support", "spam", "escalate_to_human"]
Assign each option a probability summing to 1.
Output format: {"scores": {"option": probability}}

Input: {{customer message}}

Try next: rephrase the same input 5 ways and check whether the top option stays the same.

Terminology

Related Papers

• Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)

  A KV pool smaller than the context window leaked other customers' data into 155 long-prompt responses.

• Cerebras Deploys Trillion-Parameter AI Models

  Cerebras fits Kimi K3 (3T) across 59 wafers in a pipeline, serving 1000-2000+ tokens per second

• A practical guide to AI image generation, from diffusion to agents

  Testing image APIs showed you pay per canvas, and camera-style prompts beat describing the view

• DeepInfer: High-performance inference framework for DeepSeek models

  DeepSeek open-sourced DeepInfer, its production inference engine, under Apache-2.0.

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• DeepSeek banned from U.S. federal civilian agencies under new DHS order

  DHS order bans DeepSeek from U.S. federal civilian agency IT systems

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

• DeepSeek V4.1 Flash 1017 (Max)

  DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

Related Resources

• https://shattered.io/microsoft-decision-1-qwen3-5-9b-model-2026/
• https://commandline.microsoft.com/microsoft-decision-1-model