Microsoft reportedly released a Qwen3.5-9B-based Decision-1 model
BIBIBI
AT A GLANCE
Microsoft says it released Decision-1 specifically for AI AI agents to make rapid decisions and return option probabilities.
Article
Microsoft Takes Qwen 9B turns it into a decision model, self-reported comparison GPT-6 Sol faster 35 times
Beating AI Breaking News: Microsoft has released Microsoft-Decision-1, specifically to help AI AI agents make rapid decisions. It was post-trained based on Qwen3.5-9B and only scores predefined options, returning the probability of each. For example, it can decide which model to route a user request to, or determine whether an agent should continue, retry, or hand off to a human.
Microsoft says that Decision-1 ranked first on 36 evaluation benchmarks and nearly 150000 questions. Its accuracy ranked first. It is Quyet-1.0-Large faster than the runner-up, and approximately 4.5 times faster than GPT-6 Sol.
In internal trials, the Xbox research team used it to classify more than 35 pieces of player feedback. Microsoft says its judgment quality was close to 10000 Sol, it was more than GPT-6 times faster, and its cost was approximately 14 that of the latter. The Copilot team also used it to score 1/200 responses.AI is now available on Microsoft Foundry and OpenRouter, and Vercel Decision-1 Gateway has also integrated it. It costs AI per million input token, with output free. Microsoft has not yet announced the release of 0.042 USD model weights.
Original link Decision-1 https://m.theblockbeats.info/flash/371092
Key points
01
Microsoft released Microsoft-Decision-1 to help AI AI agents make rapid decisions.
02
Decision-1 Based on Qwen3.5-9B post-training, it scores only predefined options and returns the probability of each.
03
Microsoft says that Decision-1 ranked first in 36 evaluation benchmarks and nearly 150000 questions.
04
Microsoft says that Decision-1 is Quyet-1.0-Large faster than 4.5 times faster than GPT-6 Sol, approximately 35 times.
05
Microsoft says that the Xbox research team used Decision-1 to classify more than 10000 pieces of player feedback, with judgment quality close to GPT-6 Sol, speed more than 14 times faster, and cost approximately 1/200。
06
Decision-1 is now available on Microsoft Foundry and OpenRouter, and Vercel AI Gateway has also integrated it.
07
Decision-1 Per million input token it costs 0.042 USD, with output free.
08
Microsoft has not yet announced the release of Decision-1 model weights.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Decision-1 This is not a model for directly generating long-form responses; it scores predefined options instead. For example, it can help a system select a model or decide whether to continue, retry, or hand off to a human.
Why it matters to readers
If the performance and pricing information in the article is accurate,Decision-1 could be used for model selection, feedback classification, and AI response scoring, among other tasks; however, most of the key data in the article is self-reported by Microsoft.
Beginners can think of it as a AI“quick selector” in a system: it does not provide a complete answer, but compares several options and gives their probabilities.
Risks and unknowns
The accuracy, speed, cost, and judgment-quality data in the article are primarily self-reported by Microsoft, with no independent verification found.
The article does not specify the names of the 36 evaluation benchmarks or the nearly 150000 questions, their data sources, or the testing methodology.
The article says the model is available on multiple platforms, but provides no details about actual availability.
Microsoft has not yet announced the release of the model weights.
Related Developments
Loading event timeline…
Related concepts
post-training
This term is not in the glossary yet. Browse related concepts in the glossary.
