Supertron-VL-2B: A Compact Vision-Language Model

Model Description

Supertron-VL-2B is a compact vision-language model fine-tuned from Qwen/Qwen3-VL-2B-Instruct for chart understanding, visual question answering, real-image reasoning, diagrams, document-style images, comics, and object counting.

It is designed as a smaller, faster multimodal model that can read images and answer concise questions without requiring the VRAM of larger VL models.

• Developed by: Surpem
• Model type: Vision-Language Model
• Architecture: Qwen3-VL dense multimodal transformer, 2B class
• Fine-tuned from: Qwen/Qwen3-VL-2B-Instruct
• License: Apache 2.0

---

Capabilities

Chart & Data Visualization QA

Supertron-VL-2B can answer questions about charts and plotted data, including visible values, labels, rankings, categories, and simple comparisons.

Visual Question Answering

The model can answer direct questions about real images, diagrams, document-like images, and visual scenes. It is tuned to give concise final answers rather than long explanations.

Object Counting

Supertron-VL-2B can answer object-counting questions when prompted to return a number only.

Diagrams & Structured Images

The model can interpret science-style diagrams, simple visual structures, and image-question pairs that require reading both text and visual context.

---

Get Started

from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image
import torch

model_id = "Surpem/Supertron-VL-2B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

image = Image.open("image.png").convert("RGB")
question = "What is shown in the image?"

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {
                "type": "text",
                "text": (
                    "Look at the image and answer the question concisely. "
                    "Return only the final answer, without chain-of-thought.\n"
                    f"Question: {question}"
                ),
            },
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], padding=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
generated = outputs[:, inputs["input_ids"].shape[1]:]
print(processor.batch_decode(generated, skip_special_tokens=True)[0].strip())

---

Recommended Prompts

For charts:

Read the chart image and answer the question concisely.
Return only the final answer, without chain-of-thought.
Question: <your question>

For object counting:

Look at the image and count the objects described by this phrase.
Answer with a number only, no explanation.
Phrase: <object phrase>

---

Hardware Requirements

For 4-bit quantized inference:

from transformers import BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

---

Limitations

Supertron-VL-2B is a compact 2B-class model. It may make mistakes on dense charts, small text, crowded diagrams, ambiguous labels, complex arithmetic, long documents, and tasks requiring precise segmentation masks.

---

Citation

@misc{surpem2026supertron-vl-2b,
      title={Supertron-VL-2B: Compact Vision-Language Model},
      author={Surpem},
      year={2026},
      url={https://huggingface.co/Surpem/Supertron-VL-2B},
}