vLLM Semantic Router team releases Decision 2.0 open decision models

The vLLM Semantic Router team published its Decision 2.0 family on Hugging Face at the end of September 2026. The Vega-27B repository was created on 29 September and the Lux-9B repository the same day. The release date here is the repository creation date, since the card gives no announcement date.

The Vega card describes a decision model that takes an input and a set of questions and returns a probability for each possible answer without generating text. Question types are choice, yes/no and score, and several questions about one input are answered in a single forward pass. It lists 29.37B parameters, a 32,768-token context, the Apache-2.0 licence and Qwen3.8-27B as base. It needs custom code (trust_remote_code).

The card reports JevArena 74.0, human-labelled transfer 58.7 and Jev Decision Index 56.5. It says Vega leads three same-size models on JevArena and is statistically level with AutoJev-27B. Median latency is 71.4 ms for a single-question request on one GPU. The Decision 2.0 collection holds eight items.

Secondary coverage describes six sizes from 0.6B to 27B. I could not open a primary page confirming that.

Open weights, a permissive licence, and numbers on the card. The card does not link a paper.

Sources

• https://huggingface.co/vllm-sr/Decision-2.0-Vega-27B
• https://huggingface.co/vllm-sr/Decision-2.0-Lux-9B

Editor's first pass (AI-assisted)

• Real 4: Weights are public under Apache-2.0; custom code is required.
• Useful 3.5: Built for classifying requests ahead of routing; one forward pass answers several questions.
• Notable 3.5: A routing-project team shipping a family, not a single checkpoint.
• Novel 3: Same decision format as others; the router tie-in is the angle.