d3-lite

d3-lite is the 0.8B multimodal foundation decision model of Decision 3.0, the decision models of vLLM Semantic Router. Give it an input (text or JSON, optionally with images) and the questions you need answered: pick one of several options, say yes or no, or rate on a scale. It answers them all in one call and returns a probability for every answer, without generating text.

Parameters	0.85B, including the 0.10B vision encoder
Inputs	Text or JSON, plus images (several per request)
Decision types	Choice · Yes / No · Score
License	Apache-2.0

Highlights

• Jev Decision Index 0.3, public suite: 35.65, measured with the official 0.3 kit on the released weights: all 140,178 public requests answered, none unsupported.
• +15.1 on the public suite over Decision 2.0 (its 0.8B model: 20.57 on the board), ahead in all five areas.
• Reads images: multiple images per request (PNG, JPEG or WebP), given as paths, URLs, PIL images or base64 data URLs; every question of the request sees all of them.
• Speed: a median of 13.1 ms for a text request and 49.7 ms for a request with an image, on one AMD Instinct MI325X GPU, one request at a time.
• Many questions, one call: Choice, Yes / No and Score questions about the same input are answered together, each from its own forward pass over the input, with a probability for every option.

Quickstart

pip install "transformers==5.17.0" torch torchvision pillow safetensors accelerate
pip install flash-linear-attention  # optional: fast GPU kernels for the linear-attention layers

import json

from huggingface_hub import hf_hub_download
from transformers import AutoModel

model = AutoModel.from_pretrained("vllm-sr/d3-lite", trust_remote_code=True)

# Text
result = model.system_one(
    state="The order arrived damaged yesterday. The customer has a receipt and asks for a replacement today.",
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "returns": "Refunds, replacements and damaged deliveries",
                "billing": "Payments, invoices and charges",
                "technical": "Product setup and faults"
            }
        },
        "receipt": {
            "type": "noul",
            "instructions": "Does the customer have a receipt?"
        },
        "urgency": {
            "type": "score",
            "instructions": "How urgent is this request?",
            "criteria": [
                "Routine",
                "Soon",
                "Today"
            ]
        }
    },
)
print(json.dumps(result["answers"], indent=2))

# Text and an image (or several)
receipt = hf_hub_download("vllm-sr/d3-lite", "assets/example-receipt.png")
result = model.system_one(
    state="The customer says the blender arrived cracked and attached the receipt.",
    images=[receipt],  # local paths, http(s) URLs, PIL images or base64 data URLs
    questions={
        "route": {
            "type": "choice",
            "instructions": "Which team should handle this request?",
            "criteria": {
                "returns": "Refunds, replacements and damaged deliveries",
                "billing": "Payments, invoices and charges",
                "technical": "Product setup and faults"
            }
        },
        "on_receipt": {
            "type": "noul",
            "instructions": "Does the receipt list the blender?"
        },
        "payment": {
            "type": "choice",
            "instructions": "How was the order paid?",
            "criteria": {
                "card": None,
                "cash": None,
                "gift card": None
            }
        }
    },
)
print(json.dumps(result["answers"], indent=2))

# Or as a pipeline:
# transformers.pipeline("decision", model="vllm-sr/d3-lite", trust_remote_code=True)(state=..., questions=..., images=...)

Images go before the text of the request, each read at up to 1.6 megapixels; every question of the request sees all of them.

Evaluation

Text: Jev Decision Index 0.3.1

Model	Jev Decision Index ↑	Public ↑	Same-skill tests ↑	New-domain tasks ↑
d3-lite	27.6	35.6	32.1	18.2
jiwo 0.8B	24.3	28.7	28.3	18.1
EXAONE-4.0-1.2B-JEV v0.3	23.3	30.7	26.9	16.4
OneJev 0.8B	20.5	22.2	21.8	19.9
JPT-0.8B Qwen3.5-0.8B LoRA	18.7	19.4	17.8	22.0
Decision 2.0 (0.8B)	17.5	20.6	21.3	13.9

Images: Jev Decision Index vision board 0.3.1

Model	Vision Index ↑	Public ↑	Private ↑
d3-lite	41.1	46.5	35.6
JPT-0.8B	42.1	44.6	39.7
OneJev 0.8B	41.5	46.5	36.5
Intern-Decision-0.8B	36.3	39.0	33.7

d3-lite on the public vision benchmarks († approximate rebuild):

Benchmark	d3-lite
CV-Bench	57.6
BLINK	34.8
RealWorldQA	35.4
CharXiv †	68.9
InfographicVQA †	72.1
Mind2Web †	46.0
Winoground	52.0
KIE (CORD+FUNSD) †	90.9
Moderation (Hateful Memes)	10.5
R-Bench-M	6.9
MMMU-Pro vision	9.5

_{d3-lite: internal evaluation. Others: live board data, text 2026-10-10, vision 2026-10-09.}

License

Apache-2.0 (LICENSE). Built on Qwen/Qwen3.5-0.8B (Apache-2.0).

Citation

@misc{d3_lite_2026,
  title        = {{d3-lite}: A Multimodal Foundation Decision Model},
  author       = {{vLLM Semantic Router Team}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/vllm-sr/d3-lite}}
}

Trained on AMD Instinct MI325X GPUs.