vLLM Semantic Router / KR Labs releases Vela 2.0 open-source routing and guardrail model family

vLLM Semantic Router and KR Labs released the Vela 2.0 open-source model family, with four sizes—0.3B, 0.8B, 4B, and 9B—for routing, safety checks, and span-level decisions, with weights and code released under Apache-2.0. The models receive options, labels, and scoring criteria at request time and return structured answers with probabilities and character offsets; a single call can handle routing, safety, PII, and hallucination issues simultaneously. The official statement says the 9B model achieved 0.989 AUC on prompt attack detection for unseen attack families, higher than Vela 1.0 Guard's 0.792; multilingual hate speech reached 0.855 AUC, higher than 0.646. The 0.3B model can run on CPU, with a single full routing request taking 0.09 seconds on an A40, while the 9B takes 0.60 to 0.71 seconds.