vLLM-SR Decision 3.0 (hosted)

Decision 3.0 models from vLLM Semantic Router are multimodal decision models: they do not generate chat text. You send a state (text/JSON) plus optional images and video, and a map of typed questions; the model returns calibrated probabilities (and boxes/tracks when asked).

OpenTryOn exposes the hosted API on the decide service with D3ApiAdapter and DECISION_API_KEY.

Registry id	HF model id	Size
d3-turbo	vllm-sr/d3-turbo	27B
d3-flash	vllm-sr/d3-flash	9B
d3-mini	vllm-sr/d3-mini	4B
d3-nano	vllm-sr/d3-nano	2B
d3-lite	vllm-sr/d3-lite	0.8B
d3-edge	vllm-sr/d3-edge	0.6B

On the API, vllm-sr/d3 aliases d3-turbo. Local Apache-2.0 weights: Decision 3.0 (local).

Prerequisites

# Free key (one click, no account): https://decision.vllm-sr.ai
export DECISION_API_KEY=your_key
# optional:
# export DECISION_BASE_URL=https://decision.vllm-sr.ai

Limits and model readiness: GET https://decision.vllm-sr.ai/api/info.

Question schema

Shared with Jev / Liquid d1 for noul, choice, and score. Decision 3.0 adds:

type	Purpose
box	One bounding box + confidence (top/bottom/left/right as fractions 0–1)
track	Box trajectory over a video (rate = updates per second)

Use --assets when images appear inside questions or options (placeholder map). See the d3-turbo model card for examples.

CLI

# Text routing (fashion support triage)
opentryon decide --model d3-turbo \
  --state "Customer photo: wrong size tag on linen shirt order" \
  --questions triage.json

# Product photo: where is the logo?
opentryon decide --model d3-turbo \
  --state "E-commerce packshot" \
  --image product.jpg \
  --questions box_logo.json

# Video: track a garment through a clip
opentryon decide --model d3-turbo \
  --state "Runway clip" \
  --video walk.mp4 \
  --questions track.json

opentryon decide --model d3-flash --questions q.json --state "..." --dry-run

Python

from tryon.api.vllm_sr import D3ApiAdapter

adapter = D3ApiAdapter(model_id="vllm-sr/d3-turbo")
out = adapter.decide(
    state="Red satin evening gown on mannequin",
    questions={
        "formal": {"type": "noul", "instructions": "Is this formal wear?"},
    },
    image="dress.jpg",
)
print(out["answers"])

MCP

Tools are generated from the registry, e.g. decide_d3_turbo, decide_d3_flash, … Restart MCP after registry changes. Key: DECISION_API_KEY (Connect rail: vLLM Semantic Router (Decision 3.0)).

API details

• Endpoint: POST https://decision.vllm-sr.ai/v1/systemone
• Auth: Authorization: Bearer $DECISION_API_KEY
• Body: model, questions, optional state, images, videos, assets, long_context

Compare with Typesafe Jev (text-only hosted) and Liquid d1 (local, no box/track).