ENPT

vLLM Semantic Router releases Vela 3.0: six open models for decisions plus the spans you name

Last updated:

Source · Model card: Decision 3.0 (d3)

publicado (OK do Fabricio 29/10 BRT).

Xunzhuo Liu announced Vela 3.0 on 28 Oct 2026 (09:53 BRT), a family from vLLM Semantic Router and KR Labs (collection). In one call a Vela model answers typed decisions (Choice, Yes/No, Score) and returns the exact spans you ask for: entities, relations, personal data, unsupported claims and evidence, with probabilities and character offsets. It is built for LLM routing and guardrails. All six models are Apache-2.0.

The Hugging Face collection lists vela3-fast, vela3-fast-vision, vela3-lite, vela3-mini, vela3-flash and vela3-turbo, plus a demo Space. Two designs:

  • vela3-turbo (27B) runs d3-turbo v1.0, frozen, and adds gated LoRA adapters for spans and router signals. The card says ordinary decisions come from d3-turbo bit for bit (561 of 561 identical in the vendor’s check), so its decision accuracy is d3-turbo’s.
  • vela3-fast (0.38B) is an UltraBERT-380M encoder with its own decision backbone, distilled from the d3 models’ answer distributions, English only, up to 32K tokens.

Vendor numbers (UnverifiedClaim). “#1 Decision Index ≤500M”, “beats GPT-5-mini at zero-shot classification” (vela3-turbo 81.4 vs 79.8 on the 13-set GLiFormer suite, reported numbers) and 18 ms per span question on one AMD MI325X.

What the board shows. On Decision Index 0.4.1 (data 28 Oct), vela3-fast has the best public score up to 500M parameters (50.36, vs Amber 0.4B 49.74 and GLiClass v4 edge 47.49). On Full, which adds the private tests, it is 29.27 (#170 of 243), behind GLiClass v4 edge (31.21). vela3-fast-vision (0.25B) is at Full 22.11. The board measured ECE 0.12 for vela3-fast and a 54 ms median per row.

Score. The Decision 3.0 page stays scored on d3 27B; Vela adds no new decision weights at the 27B size. The small vela3-fast is a separate catalog candidate.