vLLM Releases 'Decision 1.0', AI Models Specialized for Decision-Making

The vLLM Semantic Router team released six Decision 1.0 models and a testing environment called Decision Studio to assist application decision logic. Instead of generating text, these models provide an interface that returns the probability that each option is correct for the given state, question, and candidate criteria.

Model Composition and Features The released models consist of two lightweight encoders (Kai, Lex) and four decoders (Eos, Sol, Nox, Lux), distributed under the Apache 2.0 license on Hugging Face. The decoder series uses a Qwen3.5-based backbone, and according to the public evaluation chart in the blog post, Lux-9B achieved a weighted composite score of 76.94%, showing the highest performance among open models. (Note: The hosted model Jev scored higher at 81.05%)

Output Formats and Usage The models support three output formats.

• Choice: Probability per candidate ID (e.g., selecting between shipping/payment/technical support)
• Noul: Probability of condition satisfaction P(true) (e.g., whether it will be processed today)
• Score: Step-wise probabilities and expected score (e.g., urgency level)

Decision Studio and Deployment Decision Studio, available on Hugging Face Space, allows users to test the models directly in the browser and includes 36 English examples. Currently, a demo API based on AMD GPUs is provided, but this is for validation purposes, not for production environments. For self-hosted deployments, the TypeSafe Python SDK can be utilized. The team proposed integrating Decision into vLLM Semantic Router by default as the next step.