Vela 3.0

Vela 3.0 is the Router's own family of decision models. Every size answers the built-in routing questions on trained adapters and returns spans for PII and hallucination checks. vela3-fast is the default decision model and runs on a CPU; the larger sizes run a Decision 3.0 package with gated adapters on a GPU.

The default, which you can leave out

global:
  model_catalog:
    system:
      decision_model: {deployment: primary}  # vllm-sr/vela3-fast on the CPU

Sizes

Model	Parameters	Answers	Spans	Reads	Device memory
vela3-fast	1.52B	choice, noul, span	router	text	12 GiB
vela3-lite	0.903B	choice, noul, score, span	router	text, image, video	4 GiB
vela3-mini	4.79B	choice, noul, score, span	router	text, image, video	16 GiB
vela3-flash	8.83B	choice, noul, score, span	router	text, image, video	32 GiB
vela3-turbo	27.6B	choice, noul, score, span	router	text, image, video	72 GiB

Models are in the Vela 3.0 collection.

vela3-fast

vllm-sr/vela3-fast is a 380M UltraBERT encoder, 0.4B with the adapter a question runs on. Its ten adapters carry their own top layers and heads, so the deployment loads 1.5B parameters.

• Trained adapters answer the Router's built-in questions; choice and noul questions of your own run on its decision backbone.
• Spans run on span adapters routed by label set and wording: personal data, unsupported claims, evidence, relations and entities, with the pii and halu presets.
• It answers no score questions. A noul question is read by its instructions alone, so one that describes false or true is invalid. A set question takes one of its trained router texts.
• Precision and memory: FP32 on every device, about 7 GB once loaded; give the Router 12 GiB.
• Its answers match the package's own runtime bit for bit on a CPU (parity record).

Choose a model has its measured accuracy on the built-in signals and its calibrated thresholds.

The larger sizes

vllm-sr/vela3-lite, -mini, -flash and -turbo run a pinned Decision 3.0 package (d3-lite, d3-mini, d3-flash and d3-turbo, at the revisions their adapters were trained on) with gated LoRA adapters over it:

• The Router asks its built-in questions under the IDs these packages were trained on, so they run on trained adapters. Harm categories are asked as one hazard: noul per category, which only vela3-lite's hazard adapter answers.
• Every other choice, noul and score question answers as that d3 package does.
• They answer span questions, routed by label set and wording as vela3-fast's are. A span question reads a target of up to 32,768 tokens in windows; a longer target fails that question with scan_budget_exceeded.
• vela3-turbo answers each request on d3-turbo's state-once path first, then the router questions again with their adapter on.
• Requests with images run with every adapter off and refuse span questions.
• vela3-mini and vela3-flash run open span labels without a description by name, where their packages write a definition with the base first (parity record).

note

Only vela3-fast and the Vela 2.0 sizes have calibrated thresholds. The larger Vela 3.0 sizes use 0.5 for a built-in question without a predicate, with an uncalibrated_preset warning.