👁️ PhysicalEye-Decide-35B

Show it an image. Ask a question. Get a calibrated decision in one forward pass.

---

What it is

PhysicalEye-Decide-35B is a decision engine for images. You give it an image, a question and the possible answers. It returns a probability for every answer.

It does not generate text. It reads the input once and scores every option directly. That makes it:

• Fast: about 0.05 seconds per decision on one GPU.
• Predictable: the answer is always one of your options. No parsing, no free text.
• Calibrated: the probability is meant to be used. Set a threshold, route low-confidence cases to a person, and log every score.

Results

Test	Setting	Result
Egg inspection	Held-out test, 100 candling photos, good vs bad	98.0% accuracy, mean confidence 0.964
AI2D	200 science diagram questions, 4 options	87.0% (chance 25%)
Speed	Median latency per decision, batched server, one B200 GPU	0.051 s (p80 0.068 s)

Where to use it

Area	Example question	Type
Food and farm inspection	"Is this egg good or bad?"	choice
Factory quality control	"Is there a crack on this part?"	noul
Retail and logistics	"Which product is on the shelf?"	choice
Documents and screens	"Is the total on this receipt above 50?"	noul
Robots and edge devices	"Is the gripper holding the object?"	noul
Rating and triage	"How damaged is this package, from 0 to 10?"	score

Try these in the live demo: a real-time inspection line, a sample gallery and a playground in English, Korean and Chinese.

Quick start

pip install -U torch transformers accelerate safetensors pillow huggingface_hub
wget https://huggingface.co/masterset-ai/PhysicalEye-Decide-35B/resolve/main/decide.py

from decide import Decider

d = Decider("masterset-ai/PhysicalEye-Decide-35B")   # downloads about 69 GB on first use

result = d.decide(
    images=["egg.jpg"],
    state="Candling photo: a light is shone through the egg in a dark room.",
    question={
        "type": "choice",
        "instructions": "Is this egg good or bad?",
        "criteria": {"good": "Normal, healthy egg", "bad": "Defective, spoiled or damaged egg"},
    },
)
print(result)
# {"type": "choice", "choice": "good", "confidence": ..., "probabilities": {"good": ..., "bad": ...}}

From the command line:

python decide.py --image egg.jpg --instructions "Is this egg good or bad?" --options good bad
python decide.py --image part.jpg --instructions "Is there a crack on this part?"   # yes/no

Hardware: the weights are about 69 GB in bfloat16. Use one GPU with 80 GB or more memory (for example H100 80GB, H200, B200), or spread the model over several GPUs.

Question types

Every call takes up to 4 images, a context string (state) and one question.

Type	Fields	Returns
choice	instructions, criteria: a dict of option name to description (or null)	choice, confidence, probabilities
noul	instructions, optional criteria: {"true": ..., "false": ...}	noul: probability of yes
score	instructions, criteria: a list of level descriptions, lowest first	score (expected level), confidence, probabilities

# yes / no
d.decide(images=["part.jpg"], state="Steel bracket on a conveyor.",
         question={"type": "noul", "instructions": "Is there a visible crack?"})
# {"type": "noul", "noul": ...}

# score on a scale
d.decide(images=["box.jpg"], state="Parcel at the receiving dock.",
         question={"type": "score", "instructions": "How damaged is this package?",
                   "criteria": ["no damage", "minor dents", "torn or crushed", "destroyed"]})
# {"type": "score", "score": ..., "confidence": ..., "probabilities": {...}, "legend": {...}}

Tips

• Describe each option in plain words in criteria. Short, concrete descriptions work best.
• Put fixed context in state, such as the camera setup or the product type. Keep instructions short.
• Use the probability. Accept high-confidence decisions automatically and send the rest to a person.
• Up to 255 options per question.

How it works

1. A vision-language Mixture-of-Experts backbone reads the images, the context and the question together. It has 35B parameters in total and about 3B active per token, so it runs fast for its size.
2. A small decision readout on top of the backbone scores every option in the same forward pass.
3. A calibration temperature turns the scores into probabilities you can threshold.

Files

File	Content
model-0000X-of-00015.safetensors	Backbone weights
readout.safetensors	Decision readout
decision_config.json	Option codes, token ids and calibration temperature
decide.py	Self-contained inference code (Apache-2.0). Examples: GitHub
config.json, processor_config.json, tokenizer*, chat_template.jinja	Model, image processor and tokenizer settings

Model family

Model	Role
PhysicalEye-35B	First-person video understanding
PhysicalEye-Decide-35B	Image and text decisions with calibrated probabilities

Base model and license

• Built on FINAL-Bench/Darwin-35B-A3B-Opus, a Qwen3.5 MoE based vision-language model.
• Released under the Apache-2.0 license. The upstream Qwen license notice applies to the base weights.

Training data credits

The decision readout was trained only on openly licensed data. We thank the authors of:

• VQAv2 (CC BY 4.0, annotations) and COCO captions and instances 2017 (CC BY 4.0, annotations; images keep their Flickr licenses)
• TextVQA 0.5.1 (CC BY 4.0), ScreenQA, RICO-SCA and Screen2Words on RICO (CC BY 4.0)
• CLEVR v1.0 (CC BY 4.0), A-OKVQA (Apache-2.0), VSR (CC BY 4.0), CORD v2 (CC BY 4.0)
• Open-Jev (CC0 1.0), MMLU auxiliary train (MIT), synthetic charts made in-house
• Good and Bad Eggs Identification Image Dataset, Md Mafiul Hasan Matin, Marium Jahan, Md. Saroar Jahan, Mendeley Data, doi:10.17632/mdty358x8m.1 (CC BY 4.0)

Citation

@misc{physicaleye_decide_35b,
  title  = {PhysicalEye-Decide-35B: One-Pass Calibrated Image Decisions},
  author = {MASTERSET},
  year   = {2026},
  url    = {https://huggingface.co/masterset-ai/PhysicalEye-Decide-35B}
}