Microsoft releases Qwen-based Decision-1 scoring model

Microsoft launched a Qwen3.5-9B decision model claiming strong benchmark accuracy, much lower latency, and free output tokens.

Microsoft has released a Qwen3.5-9B-based decision model, called Decision-1 by The Decoder and Microsoft-Decision-1 by Help Net Security, for classifications, evaluations, routing, scoring, and potentially guiding agents. The later report says it was post-trained for single-pass scoring and returns a calibrated probability for yes/no, multiple-choice, rating, and rubric tasks rather than chat. Across 36 benchmarks covering nearly 150,000 questions, Microsoft reported leading results; The Decoder gives 83.5% accuracy at 85 ms and a lead over Jev 1.13.0, while Help Net Security describes the set as blind and does not repeat those figures. Both sources say it was about 2.5 times faster than H2O-Lightning-4B. Help Net Security adds roughly 35 times faster median latency than GPT-6 Sol and cites internal Xbox, Copilot, incident-response, and Microsoft Discovery tests; The Decoder notes Cloudflare’s Qwen-based Clef models were left out of the comparison. The model is on Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with output tokens free.

• Microsoft released a decision model named Decision-1 in one report and Microsoft-Decision-1 in another, based on or post-trained from Qwen3.5-9B.
• It is aimed at structured decisions—classifications, evaluations, routing, yes/no, multiple-choice, ratings, and rubric grading—and may guide agents, rather than general chat.
• Microsoft said it led 36 benchmarks covering nearly 150,000 questions; one report calls them blind and says the model returns a calibrated probability per option.
• One report cites 83.5% accuracy at 85 ms latency and a lead over Jev 1.13.0; both say it was about 2.5 times faster than H2O-Lightning-4B.
• A later report adds that median latency was roughly 35 times faster than GPT-6 Sol; the earlier report says Cloudflare’s Qwen-based Clef models were excluded from the comparison.
• It is available on Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with output tokens free.
• Internal tests cited in the later report covered Xbox, Copilot, incident response, Microsoft Discovery, feedback sorting, grading, and agent replanning, with competitive quality at lower latency and cost.