This website uses cookies

Read our Privacy policy and Terms of use for more information.

CHINA AI2X BRIEFING

How AI is reshaping China’s Industries


➤ Must Read

_________

Why China’s AI firms are betting on decision models

From Alibaba to startups in Hangzhou and Shanghai, decision models find their place in software and machinery. Microsoft builds on Qwen.

Published on Nov. 10, 2026

Infographic: Asia Waypoint

A drone operator describes a target, and a compact AI assists in locating it. Hangzhou startup OmAI already provides this for aerial inspections. The Chinese robot makers Unitree and Deep Robotics have started deploying the startup's decision models, reported the tech portal 36Kr. OmAI's CEO sees the sudden enthusiasm for decision models as recognition of a business he has pursued for years (in Chinese).

“The emergence of decision models was extremely exciting for our entire team,” Zhao Tiancheng, OmAI’s chief executive and chief scientist, told 36Kr. Much of its work over the past five years involved similar functions. The new category “decision models” suddenly gave them a shared name.

A decision model picks from predefined choices and returns a score or probability. Customer-service systems can use this to route tickets. AI agents use it to choose tools. The result feeds directly into workflows without needing to extract answers from generated text.

“It must be fast, low-cost and accessible directly on the device,” Zhao said of the immediately usable judgments his clients need. The appeal is clear when workflows demand many such judgments in sequence.

Jev, a decision model launched by the pioneering San Francisco startup TypeSafeAI in mid-September, alerted the whole industry to the power of simple decisions. It cannot generate text. It just provides answers like yes or no, option numbers, or probabilities. It does this with super fast response times between 70 and 500 milliseconds, however. OpenAI, Amazon, Cloudflare, and Microsoft have since joined the field and launched their own decision models.

“Large language models can actually make many types of decisions as well,” Zhao said. The beauty of a specialized model is being able to optimize a narrow task. In a Microsoft demo cited by 36Kr, a decision model completed a request in 182 milliseconds, compared to 6.4 seconds for a language model.

StartLux targets this market from Shanghai. It released StartLux-Prism, an open-source engine for decision models, Chinese media reported on November 5. The engine aims to make them faster, cheaper, and easier to run locally.

Chuan-Wei Kuo, StartLux’s co-founder and CTO, sees uses in customer-service routing, web operations, and AI agent code execution. One StartLux model ranked first among 174 in a Hugging Face decision-model ranking, the company says.

“A structured decision model designed for high-frequency business decisions” is how Alibaba Cloud describes its own offering. Its preview model handles classification, yes-or-no decisions, and scoring, providing probabilities and confidence scores. Alibaba lists applications like ticket routing, content moderation, and agent routing.

For business buyers, such decisions recur throughout the day. Running them cheaply and quickly could make automation viable across more processes. Local deployment also attracts customers hoping to keep sensitive data on their own systems.

A model from Alibaba served as the foundation for Microsoft's new decision model. When Microsoft launched Decision-1 on October 9, it said that it had adapted Qwen3.5-9B for decision scoring.

“To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring,” the company wrote. The company plans to introduce versions based on its own MAI models and OpenAI models later.

Decision-1 costs US$0.042 per million input tokens with free output, exactly what TypeSafe charges for Jev, and it arrived the same day TypeSafe announced a big investment round.

Microsoft uses its Decision-1 internally to sort over 10,000 pieces of player feedback for Xbox Research. The Copilot team used it for grading chat and agent responses, “The New Stack” reported on October 10.

For Alibaba, the significance extends beyond selling its decision service, as Qwen often underpins competing products. The contest involves both the model and the specialized service customers use.

“No answer token is sampled, no free-form answer is parsed, and no chain of thought is generated,” states an October research paper on Chinese Decision from the Chinese Academy of Sciences’ Institute of Computing Technology. Its approach involves extracting probabilities for available answers in a single model pass.

The researchers adapted a Qwen3.5 model with 800 million parameters to cover general questions and fields like medicine, law, and finance. A larger model did not surpass the smaller one's average performance in these areas.

A service choosing the right customer-support department has a defined set of answers. Things get more complicated with physical AI, however. A robot must interpret changing camera feeds, movements, and instructions. It must also anticipate future actions, as some quick responses are only effective in advance. Imagine a wall coming up. The decision to avoid it cannot be made when the robot hits it.

“These are precisely the typical problems encountered when purely cloud-based large language models are used in physical-world applications,” Zhao told 36Kr. Decision models are useful in edge devices. OmAI’s approach combines small models running on devices with continuous data processing. Not every decision must travel to the cloud.

The startup's VLX-Seek model operates on the OttoPlex platform to locate targets and areas described in natural language. Results assist drone operators and task-management systems. Other models interpret video, reconstruct events, and guide navigation, offering more than single text-based choices like ticket routing.

Speed is a major incentive here. The company claims local processing achieves responses around 0.1 seconds, compared to over five seconds in the cloud. On-device processing also helps when network connections are unavailable or unstable or when sensitive data must stay local.

Zhao anticipates that in the future, reasoning and immediate action will split between cloud and device. “The slow-thinking processes will run in the cloud, while all fast thinking and rapid decision-making will take place directly on the device,” he said. “This is already how we operate many of our products.”