vLLM Releases Six Open-Source Decision Models Built on Qwen

On October 10, 2026, the open-source inference platform vLLM released six decision models, following a similar approach to Microsoft's Decision-1 by post-training different versions of the Qwen model. The models are designed for high-frequency, small-scale judgment calls in agent workflows.

Unlike general-purpose LLMs, decision models do not chat, write code, or summarize documents. Instead, they take a context and a set of predefined candidate answers, then output calibrated probabilities for each option in a single forward pass. These outputs are consumed by programs, not humans.

According to vLLM's Decision 2.0 hub, the models can handle request classification, model routing, search relevance judgment, scoring AI replies or agent actions, safety blocking, data validation, ticket triage, selecting the next step from fixed actions, and large-scale text labeling. A demo project showcases LLM routing: given a prompt like "What is 17 x 24?", the model returns decisions across six dimensions, including model choice, complexity, whether to attempt jailbreak, required tools, fact-checking needs, and latency priority.

The release adds to a growing wave of decision models, most of which are built on different versions of Qwen, while OpenAI's Decisions API and TypeSafe's Jev are notable exceptions. This type of model is likely to remain as an infrastructure layer for agents rather than replace general-purpose LLMs.