Microsoft-Decision-1 lands in Foundry: 85ms scoring, free output
The fastest decision model yet beats Jev 2.8x on speed at $0.042/M in — plus a new Decisions API to learn, and Step 5's open weights drop Oct 15.
Copy markdown
A decision model, not a chatbot — and the fastest one yet
Microsoft-Decision-1 scores a fixed set of options with a calibrated probability instead of writing text, hitting 83.5% accuracy at an 85ms median (p95 125ms) on Microsoft's own 36-benchmark suite — about 2.8x faster than TypeSafe's Jev at the same $0.042/M input, with output billed free. It's in public preview in Microsoft Foundry and on OpenRouter via Azure, so you can wire it into routing, classification or agent guardrails today.
The catch: it speaks a Decisions API, not chat completions
You call it through the Decisions endpoint — send content, get back a probability per option — so your existing chat-completions SDK won't reach it and you'll rewrite that call. Context caps at ~33K tokens and input is text-only for now.
Built on Qwen3.5-9B, with calibration still on Jev's side
The weights are Alibaba's open Qwen3.5-9B post-trained by Microsoft; Microsoft says later versions will rebase on its own MAI models and OpenAI's, no date given. Jev stays better-calibrated (93.7 vs 92.2 in Microsoft's own table), so test the probabilities on your own data before you automate decisions on them.
Update: Step 5's open weights land Oct 15
StepFun's 600B-total / 27B-active, 1M-context Step 5 Preview — already live on OpenRouter at $1/M in — gets its open-weight drop on Oct 15. The official Hugging Face repo is still private and an unverified mirror has already surfaced, so wait for the real release rather than pulling the leak.