Vela 3.0 introduces six open models for routing decisions and span extraction

What happened

vLLM highlights Vela 3.0 from the vLLM Semantic Router, KR Labs, and AMD teams, combining routing decisions and span extraction in one call with six open models from 0.3B to 27B. In the quoted announcement, @XunzhuoLiu reports that the 380M model ranks #1 on the Decision Index for models ≤500M, beats GPT-5-mini at zero-shot classification, and takes 18 ms per span question.

Read next

• More changes in today’s briefingGoogle DeepMind announces Genie 4 for real-time world simulation with native audio

The story so far

The vLLM Semantic Router, KR Labs, and AMD teams introduced Vela 3.0, which vLLM describes as six open models ranging from 0.3B to 27B parameters that combine routing decisions and extraction of user-specified text spans in one call. In the announcement quoted by vLLM, @XunzhuoLiu said the 380M model ranks first on the Decision Index among models with 500M parameters or fewer, beats GPT-5-mini at zero-shot classification, and takes 18 milliseconds per span question.