Decision-2.0-Vega-27B

Decision-2.0-Vega-27B is the 27B model of Decision 2.0, the decision models of vLLM Semantic Router. Give it an input (text or JSON) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all at once and returns a probability for every answer, without generating text.

Highlights

• Top JevArena score of its size: 74.0 among the 4 same-size models compared, statistically level with AutoJev-27B (72.1).
• The strongest Decision 2.0 model: +11.2 on the Jev Decision Index over Decision-2.0-Lux-9B.
• Speed: a median of 71.4 ms per single-question request on a single GPU.
• Many questions, one pass: Choice, Yes / No and Score questions about the same input are answered together in one forward pass, with a probability for every option.

Quickstart

pip install "transformers>=5.17" torch safetensors peft

import json

from transformers import AutoModel

model = AutoModel.from_pretrained("vllm-sr/Decision-2.0-Vega-27B", trust_remote_code=True)
result = model.system_one(
    state="The order arrived damaged yesterday. The customer has a receipt and asks for a replacement today.",
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "returns": "Refunds, replacements and damaged deliveries",
                "billing": "Payments, invoices and charges",
                "technical": "Product setup and faults"
            }
        },
        "receipt": {
            "type": "noul",
            "instructions": "Does the customer have a receipt?"
        },
        "urgency": {
            "type": "score",
            "instructions": "How urgent is this request?",
            "criteria": [
                "Routine",
                "Soon",
                "Today"
            ]
        }
    },
)
print(json.dumps(result["answers"], indent=2))

# Or as a pipeline:
# transformers.pipeline("decision", model="vllm-sr/Decision-2.0-Vega-27B", trust_remote_code=True)(state=..., questions=...)

Evaluation

JevArena

_{Every model answers the same frozen prompts, scored the same way; missing or invalid answers count as errors. Human-labelled transfer is the median macro-F1 over 15 human-labelled tasks (×100).}

Jev Decision Index

_{Decision 2.0: independent reproduction with the official 0.2.1 kit on the released weights; others: public board snapshot, 2026-09-28. Training data audited at row level against all Index test items.}

License

Apache-2.0 (LICENSE).

Citation

@misc{decision_2_0_vega_27b_2026,
  title        = {{Decision-2.0-Vega-27B}: A Decision 2.0 Model for Structured Decisions},
  author       = {{vLLM Semantic Router Team}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/vllm-sr/Decision-2.0-Vega-27B}}
}