Microsoft Launches Decision-1 Model Built on China's Qwen3.5-9B

Microsoft is entering the growing category of so-called decision models. Its Microsoft-Decision-1 is built on the open Qwen3.5-9B model created by China's Alibaba Cloud, not on the Redmond company's own weights.

What Microsoft announced

The launch was announced in a Friday blog post by Achint Srivastava, VP of software engineering in the Office of the CTO at Microsoft. The model is available in Microsoft Foundry, and availability on OpenRouter has been billed as imminent.

The Register reports that Microsoft, without giving reasons, said it will soon rebase Decision-1 on its own models and those from OpenAI.

Decision models are quickly emerging as an important new category in AI. - Achint Srivastava, VP of software engineering in the Office of the CTO at Microsoft

What a decision model is

According to The Register, the race was kicked off by Jev from TypeSafe AI, announced about three weeks earlier. It is an LLM tuned to answer certain types of questions with a limited range of responses, rated by probability.

Such models are meant for business uses where open-ended text of uncertain accuracy is undesirable. Their strengths are speed, low cost and constrained output. One of Jev's selling points is that decision models don't hallucinate the way standard LLMs do. The Register notes, however, that they can still make errors, and researchers are already exploring how those errors might be magnified.

Among competitors, The Register lists OpenAI's Decisions API (in public beta), Cloudflare's Clef, Strands Decider 2B, Liquid AI's d1, Perplexity's Decisions API, a Snowflake model, Surogate Rune, and H2O.ai's H2O-Lightning-4B, also Qwen-based, built on the 4B version. The outlet writes that more than 100 such models are now competing.

The company's numbers

Microsoft claims that in its latency test Decision-1 is 2.5x faster than H2O-Lightning-4B and 2.8x faster than Jev. It also reports the highest accuracy, 83.5 percent across 36 benchmarks, and second place in confidence score (92.2 percent), behind Quyet-1.0-Large.

In text classification, the model is said to be more than 20x cheaper than OpenAI's GPT-6 Sol. These are the vendor's claims, not independent measurements.

What it means for businesses

For teams that classify tickets or moderate content, a small model with a fixed set of responses could be a cheaper alternative to a large LLM. The choice of an Alibaba Cloud base may raise questions at companies with strict requirements on model provenance, though the announced move to Microsoft and OpenAI models suggests the vendor treats the current version as transitional.