Microsoft Corporation's latest decision-making AI model anchors its underlying architecture on Alibaba Group Holding Ltd (NYSE: BABA)'s open-source model.

Microsoft officially launched its brand-new decision model, Microsoft-Decision-1, on October 12, built through post-training on Alibaba's Qwen3.5-9B. Microsoft stated that the model is trained on top of Qwen3.5-9B, and in structured decision-making tasks, it surpasses popular decision models such as the LLM and Jev 1.13.0 in terms of latency and accuracy.

Microsoft's Xbox research team has already used it to process over 10,000 pieces of player feedback, and Microsoft Corporation (NASDAQ: MSFT) CEO Satya Nadella also revealed on social media that, based on the new model's outstanding performance, Microsoft is testing and applying the new model across all areas, including incident response, quality control, and scientific discovery.

Unlike traditional large models that generate natural language, decision models focus on outputting predicted values and probabilities. Their smaller size and faster response speed precisely meet enterprise-end urgent demand for low cost and high efficiency in frequent judgment tasks. This sends a clear signal to investors: lightweight, specialized decision models are unleashing enormous commercial value.

It is understood that in September, the explosive popularity of the Jev model sparked a wave of decision models. This type of model does not generate natural language but outputs predicted values, probabilities, and confidence levels. It is far smaller than large language models (LLMs) and responds faster, exactly meeting the massive demand for frequent judgments in AI tasks.

Statistics show that among 129 open-source decision model projects on HuggingFace, 84 were developed based on Alibaba's Qwen, accounting for more than 65%, including well-known decision models launched by companies such as Microsoft, Perplexity, Cloudflare, and H2O.ai, all of which adopt Alibaba's Qwen.

Currently, Microsoft has conducted extensive testing of the model across the entire company, covering multiple core business areas from incident response and quality control to scientific discovery. With the model fully launched on Microsoft Foundry and the OpenRouter platform, decision models based on the Qwen architecture are expected to reshape the workflows and investment logic of enterprise-grade AI applications.

New model shows significant performance and cost advantages

Microsoft's official introduction emphasized the performance breakthroughs of the new model based on the Qwen architecture.

Achint Srivastava, vice president of software engineering in Microsoft's CTO office, wrote that Microsoft-Decision-1 achieved the highest accuracy across 36 benchmark tests covering nearly 150,000 questions.

In terms of speed, the model's performance is especially prominent. Test data show that its measurement speed is 2.5 times that of the runner-up model H2O-Lightning-4B v1.1, and 35 times that of GPT-6 Sol.

In addition, the model also performs stably in robustness and safety. When option descriptions are paraphrased or shuffled, its decision change rate averages only 1.3%.

In terms of pricing, the model's input token cost is US$0.042 per million tokens, while output tokens are completely free, greatly lowering the threshold for enterprise-grade use.

Comprehensive internal testing accelerates enterprise-grade adoption

Outstanding underlying performance prompted Microsoft to accelerate the internal application conversion of the model. Satya Nadella emphasized that Microsoft is testing and applying this new model in all areas to verify its reliability in complex environments.

In data annotation, Microsoft's Xbox research team has used the model to process more than 10,000 pieces of open-ended player feedback from questionnaires and social media.

The results show that its quality is comparable to GPT-6 Sol, but it runs 14 times faster and costs 200 times less. In the field of quality control, tests by the Copilot team show that the new model is equally competitive with GPT5.6 Luna and is 100 times faster.

In addition, in adaptive replanning tasks for scientific discovery, the model's consistency is 46 times that of LLM-based scoring, with speed increased by nearly 4 times. These intensive internal test results indicate that decision models based on the Qwen foundation are effectively being converted into core tools for improving enterprise operational efficiency.