vLLM Semantic Router, KR Labs, and AMD Release Vela 3.0 Family of Six Open Models for Single Call Routing and Span Extraction
Introducing Vela 3.0 📡 Name it. Vela finds it. Six open models (0.3B–27B): decisions + the spans you name. One call. ⚡ 380M: #1 Decision Index ≤500M 🎯 Beats GPT-5-mini at zero-shot classification 🔍 18 ms per span question 🤗 t.co/F1fYdiSNzX #Jev #Decison
Context
As multi model agent pipelines rely on fast request routing, safety checks, and entity extraction before invoking expensive frontier LLMs, Vela 3.0 collapses zero shot classification and span extraction into a single sub 20 millisecond pass across models as small as 0.3B parameters.
The vLLM Semantic Router project builds open source routing infrastructure to direct prompts across heterogeneous language models and guardrails. On Wednesday morning, lead maintainer Xunzhuo Liu and the official vLLM account announced Vela 3.0, developed in collaboration with KR Labs and AMD and released as an open weight collection on Hugging Face.
Deep dive
The vLLM Semantic Router team, in partnership with KR Labs and AMD, released Vela 3.0 on Wednesday morning, an open weight family of six specialized models ranging from 0.3 billion to 27 billion parameters designed to perform both routing decisions and named span extraction in a single inference call. The collection is available immediately on Hugging Face alongside an interactive browser studio running on CPU.
The six released checkpoints span text and vision routing tiers: vela3 fast (a 0.4B UltraBERT encoder), vela3 fast vision (a 0.3B multimodal router that handles image classification, span localization, and optical character recognition checks at 0.25B active parameters per call), vela3 lite, vela3 mini, vela3 flash, and vela3 turbo up to 27B parameters. Rather than chaining separate classifier, guardrail, and named entity recognition models in front of an LLM, developers can pass arbitrary category labels and span questions to Vela 3.0 and receive both the routing decision and the exact extracted text or image spans in one pass.
According to the release benchmarks shared by Xunzhuo Liu and the vLLM project, the 380 million parameter encoder ranks number one on the public Decision Index for models under 500 million parameters, outperforms OpenAI's GPT 5 mini on zero shot classification, and answers span extraction queries in roughly 18 milliseconds.
What to expect next
Production agent stacks increasingly spend significant latency and GPU budget on pre routing, personal data redaction, and tool selection before the main reasoning model ever runs. By open sourcing six sub 500M to 27B text and vision routers that return decisions and named spans in roughly 18 milliseconds and run on standard CPUs, the vLLM, KR Labs, and AMD collaboration makes semantic routing practical for every self hosted deployment.