Vela-2.0-4B

A 4B Vela 2.0 decoder, ahead of the Vela 1.0 specialists on every built-in signal but fact check, on a GPU.

Property	Value
Repository	vllm-sr/Vela-2.0-4B, pinned at c1e64d4f872c
Family	Vela 2.0
Parameters	4.21B loaded
Backbone	Qwen3.5 text
Precision (--dtype auto)	BF16 on a GPU, FP32 on a CPU
Weights	about 8.4 GB at BF16, 16.9 GB at FP32
Device memory	16 GiB or more
Reads	text

Question types

Type	Answers
choice	Yes
noul	Yes
score	Yes
set	Yes
span	Yes, on the router, broad span heads

Built-in questions

The Vela 2.0 family was trained on the Router's built-in questions, so the Router asks each one in its trained form; the support matrix lists which.

A rule without a predicate takes this model's calibrated threshold:

Preset	Threshold
jailbreak (jailbreak)	0.63
domain (domain)	0.45
pii (pii)	0.05
fact check (fact_check)	0.9984
user feedback (feedback)	0.33

Hardware

Each time it loads, the model compares its answers with the reference answers recorded for its revision. They are recorded for:

• CPU (cpu)
• NVIDIA Hopper (sm_90) (cuda:sm_90)
• AMD GPU (ROCm) (rocm)

Where a hardware reference and its device class's are both recorded, the hardware one is checked.

Choose a model lists the validated hardware.

Quantization

The runtime loads no quantized packages of the Vela 2.0 family: it runs at its published precision.

Deploy it

Serve it on its own in Engine mode, to ask it directly:

CPU

vllm-sr serve vllm-sr/Vela-2.0-4B --engine --platform cpu

Serve it on a CPU from the CPU image, as --platform cpu does: the ROCm image's PyTorch has no CPU LAPACK, which this backbone needs there.

AMD GPU

vllm-sr serve vllm-sr/Vela-2.0-4B --engine --platform rocm

NVIDIA GPU

vllm-sr serve vllm-sr/Vela-2.0-4B --engine --platform cuda

Or make it the Router's decision model:

config.yaml

global:
  model_catalog:
    deployments:
      primary:
        provider: model_runtime
        artifact: vllm-sr/Vela-2.0-4B
        device: auto
    system:
      decision_model: {deployment: primary}

device: auto takes the first validated GPU with enough free memory, else the CPU; name rocm:0 or cuda:0 to pin a GPU.

To run it on another host and attach the Router to it:

vllm-srun serve vllm-sr/Vela-2.0-4B --device auto --host 0.0.0.0 --port 8100

Then set endpoint: http://<host>:8100 and model: Vela-2.0-4B on the deployment; see Attach to a runtime you run.

Learn more

• Model card on Hugging Face
• vela2-decision-model-sizes record
• vela2-parity record
• cuda-vela2-performance record
• Vela 2.0: the family's sizes and how they differ