Decision 3.0: The State-of-the-Art Foundation Multimodal Decision Models. Free.

  • #1 Text 66.26
  • #1 Vision 71.26
  • 6 sizes, 0.6B to 27B

One call. Every answer.

Send text, an image or a video with your questions. Get a probability for every answer.

What happens in this clip?

  • liftoff0.9942
  • on the pad0.0053
  • landing0.0005
Request
{  "model": "vllm-sr/d3",  "state": "A clip filmed at a launch site.",  "videos": ["data:video/mp4;base64,…"],  "questions": {    "event": {      "type": "choice",      "instructions": "What happens in this clip?",      "criteria": {        "liftoff": "The rocket lifts off and climbs into the sky",        "on the pad": "The rocket stays on the launch pad",        "landing": "The rocket comes down and lands"      }    }  }}
Layered buttes and pale rock outcrops on Mount Sharp, Mars, under a hazy sky. Image: NASA/JPL-Caltech/MSSS

Which part of the scene is closest to the camera?

  • dark hills0.9944
  • pale rocks0.0037
  • ridge0.0019
Request
{  "model": "vllm-sr/d3",  "state": "A photo taken by a rover on Mars.",  "images": ["data:image/jpeg;base64,…"],  "questions": {    "nearest": {      "type": "choice",      "instructions": "Which part of the scene is closest to the camera?",      "criteria": {        "dark hills": "The dark, rounded hills",        "pale rocks": "The pale, layered rocks",        "ridge": "The flat-topped ridge on the horizon"      }    }  }}
Our production app has returned 500 errors since the 09:00 deploy. Customers cannot check out. Please help now.

Which team should handle this ticket?

  • billing0.0006
  • engineering0.9994
  • sales0.0000

How urgent is this ticket?

  • Low0.0003
  • Normal0.0009
  • Urgent0.9988
Request
{  "model": "vllm-sr/d3",  "state": "Our production app has returned 500 errors since the 09:00 deploy. Customers cannot check out. Please help now.",  "questions": {    "route": {      "type": "choice",      "instructions": "Which team should handle this ticket?",      "criteria": {        "billing": "Payments, invoices and refunds",        "engineering": "Outages, errors and bugs",        "sales": "Plans, pricing and contracts"      }    },    "urgency": {      "type": "score",      "instructions": "How urgent is this ticket?",      "criteria": ["Low", "Normal", "Urgent"]    }  }}

Real answers of vllm-sr/d3, recorded 2026-10-13, rounded to 4 decimals. Read the docs

#1 on text and vision.

The Pareto frontier at every size on vision, five of six on text.

Text board, d3
66.26
Vision board, d3
71.26
Sizes on the vision frontier
6/6
Sizes on the text frontier
5/6
Vision board of the official Jev Decision Index 0.3.2: 33 entries by parameter count on a log scale against the official Full score. The six d3 models, d3 27B at 71.26 down to d3-edge 0.6B at 31.71, are exactly the Pareto frontier.
All six d3 sizes form the vision Pareto frontier.
Text board of the official Jev Decision Index 0.3.2: 147 entries by parameter count on a log scale against the official Full score. d3 27B is the highest-scoring entry at 66.26; d3, d3-flash, d3-mini, d3-nano and d3-lite lie on the Pareto frontier; d3-edge 0.6B is just off it, #2 in its size band.
Five of six d3 sizes are on the text Pareto frontier; d3-edge is #2 in its size band.

d3 models Other entries Pareto frontier Parameters (log scale) against the official Full score, 2026-10-12. A size band is 0.75x to 1.33x a model’s parameters. Official Jev Decision Index

Six sizes. One family.

Every size reads text, images and video.

Decision 3.0 models, their sizes, official Full scores and Hugging Face model cards
ModelSizeTextVisionModel card
d327B66.2671.26Hugging Face
d3-flash9B55.7966.89Hugging Face
d3-mini4B52.6163.16Hugging Face
d3-nano2B38.8654.67Hugging Face
d3-lite0.8B32.3342.11Hugging Face
d3-edge0.6B20.2431.71Hugging Face

Official Full scores on the Jev Decision Index 0.3.2 text and vision boards (2026-10-12). Use a model as vllm-sr/<name>.

All models on Hugging Face

Sponsored by

AMD

Runs on AMD Instinct MI325X GPUs.


Acknowledgements

Training

Trained with PyTorch on AMD Instinct GPUs by the AMD vLLM-SR team.

Evaluation

Evaluated on the Jev Decision Index on Hugging Face.