madrisight/madrimed1.2-VL-2B

MadriMed 1.2

2B Multimodal Medical Vision-Language Model

---

Overview

MadriMed 1.2 is the second compact multimodal medical vision-language model by Madrisight. Building upon madriMed-VL-2B, MadriMed 1.2 delivers significant performance leaps across medical image understanding (Radiology, Pathology, CT, MRI, X-Ray) and medical reasoning.

It Delivers robust clinical reasoning capabilities, drastically reducing hallucinations and diagnostic contradictions while setting a new state-of-the-art benchmark for 2B medical vision models.

What's New in MadriMed 1.2?

1. Massive Diagnostic Accuracy Gains:
   • Massive jump on VQA-RAD overall score and most of the other evalution benchmarks.
2. Resistant to Generation Hallucinations:
   • Substantially reduced token corruption, repetition and positive-bias tendencies observed in version one.
3. Structured Formating:
   • Trained with graduated correctness scoring and reasoning-integrity penalties to enforce structured.

---

Quick Start

Installation

pip install torch torchvision transformers

Input

prompt = """Question: Examine the mammogram image shown above. Which of the following findings is most evident?

Answer Choices:
A. Well-circumscribed round mass with benign features
B. Clustered microcalcifications within an area of irregular density
C. Fat-containing lesion consistent with lipoma
D. Diffuse bilateral breast edema

Follow this exact step-by-step reasoning structure.

### 1. Key Clinical Findings
Extract the patient's age, gender, main symptoms, duration, objective physical exam findings, and abnormal laboratory/imaging results from the provided question and the images.

### 2. Differential Diagnosis
List 2-3 highly probable conditions based on the extracted findings. Briefly state why the current presentation fits.

### 3. Option Elimination 
Analyze each provided choice. Write one sentence explaining why it is either correct or incorrect based on clinical guidelines.

### 4. Conclusion & Final Answer
State the final diagnostic/management conclusion.

### 5. Final correct choice
Provide final single correct choice.
"""

Input Image A

Input Image B

Run the model directly

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
from PIL import Image

MODEL_ID = "madrisight/madrimed1.2-VL-2B"
device = "mps" if torch.backends.mps.is_available() else "cpu"

model = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID,
    device_map=device,
    dtype=torch.float32,
    trust_remote_code=True,
).eval()

processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)

# Multi-image input
image_paths = ["images/MM-1-a.png", "images/MM-1-b.png"]
images = [Image.open(path).convert("RGB") for path in image_paths]

# Chat Message Payload
messages = [
    {
        "role": "system",
        "content": "You are an expert clinical medical educator and USMLE board reviewer.",
    },
    {
        "role": "user",
        "content": [{"type": "image"} for _ in images] + [{"type": "text", "text": prompt}]
    }
]

# Tokenize, Process, and Generate
with torch.inference_mode():
    text = processor.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    inputs = processor(
        text=text,
        images=images,  # Pass the flat list of PIL Image objects
        return_tensors="pt",
        padding=True,
        truncation=True,
    ).to(device)

    generated_ids = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=False,
    )

    output_text = processor.batch_decode(
        generated_ids[:, inputs.input_ids.shape[1]:],
        skip_special_tokens=True
    )[0].strip()

Output

### 1. Key Clinical Findings
- **Patient's age and gender**: Not explicitly mentioned, but mammograms are typically used for women.
- **Main symptoms**: Not specified, but mammograms are often performed for breast abnormalities.
- **Duration**: Not specified, but mammograms are usually performed at the time of suspicion.
- **Objective physical exam findings**: Not specified, but mammograms are used to detect abnormalities.
- **Abnormal laboratory/imaging results**: Not specified, but mammograms are used to detect breast abnormalities.

### 2. Differential Diagnosis
- **Well-circumscribed round mass with benign features**: This could be a benign lesion such as a fibroadenoma.
- **Clustered microcalcifications within an area of irregular density**: This could indicate a malignant process, such as ductal carcinoma in situ (DCIS) or invasive breast cancer.
- **Fat-containing lesion consistent with lipoma**: This is a benign lesion and could be a lipoma.

### 3. Option Elimination
- **Well-circumscribed round mass with benign features**: This is a plausible finding, but it is less likely to be the primary concern in a mammogram without additional symptoms or findings.
- **Clustered microcalcifications within an area of irregular density**: This is a highly suspicious finding for malignancy, as it is often associated with breast cancer.
- **Fat-containing lesion consistent with lipoma**: This is a benign finding and less likely to be the primary concern in a mammogram without additional symptoms or findings.

### 4. Conclusion & Final Answer
The mammogram findings are most indicative of a malignant process, specifically breast cancer, due to the presence of clustered microcalcifications within an area of irregular density.

### 5. Final correct choice
B. Clustered microcalcifications within an area of irregular density

---

Benchmark Results

MadriMed 1.2 was rigorously evaluated across 7 medical image and text benchmarks.

🔬 Multimodal & Visual Question Answering

Benchmark	MediX-R1 (2B)	Qwen3.5-2B-MedVL (OpenMed)	MadriMed 1.2 (2B)
VQA-RAD	53.9%	55.21%	64.30%
PathVQA	42.8%	47.48%	48.29%
SLAKE (Radiology)	65.4%	68.80%	69.36%

VQA Comparison with 4B

Benchmark	Metric	MediX-R1 2B	Qwen3.5-2B-MedVL	MedMO-4B-Next	MedGemma 1.5 4B	MedGemma 4B	MadriMed 1.2 2B
SLAKE	Overall	65.40%	68.80%	74.00%	59.80% Tok-F1	72.30% Tok-F1	69.36%
↳ Closed	Accuracy	68.40%	N/R	78.00%	82.80%	87.60%	71.83%
↳ Open	—	57.60%	N/R	—	59.70% Tok-F1	—	68.13%
VQA-RAD	Overall	53.90%	55.21%	59.60%	48.10% Tok-F1	49.90% Tok-F1	64.30%
↳ Closed	Accuracy	61.30%	N/R	79.70%	70.20%	69.10%	74.90%
↳ Open	—	45.10%	N/R	—	—	—	51.00%
PathVQA	Overall	42.80%	47.48%	73.30%	—	—	48.29%
MedXpertQA (MM+Text)	Accuracy	—	—	—	20.90%	18.80%	18.02%

Benchmark	Metric	MedMO-4B-Next	MedGemma (4B)	MadriMed-VL-2B (v1)	MadriMed 1.2 (2B)
MedXpertQA (Multimodal)	Accuracy	27.0%	24.43%	21.50%	23.50%
MedXpertQA (Text)	Accuracy	16.5%	14.2%	11.02%	13.55%

---

📝 Text only Benchmarks

Benchmark	MedGemma 1.5 (4B)	MedGemma (4B)	MediX-R1 (2B)	MadriMed 1.2 (2B)
PubMedQA	68.2%	73.4%	47.2%	66.6%
MedMCQA	55.7%	59.8%	49.2%	43.1%
MedQA-4opts	69.1%	64.4%	49.7%	41.56%
MedXpertQA (Text)		14.2%		13.55%

---

SLAKE Per-Category Performance

Category	EM	Total VQA
Abnormality	40.4%	52
Color	81.8%	33
KG	57.8%	109
Modality	91.8%	85
Organ	71.7%	99
Plane	90.9%	33
Position	61.4%	171
Quantity	73.1%	52
Shape	42.9%	7
Size	69.2%	65

---

Medmcqa (Text-Only) Per-Category Performance

By Choice Type

Choice Type	Accuracy	Correct / Total
Multi	41.62%	569 / 1,367
Single	43.82%	1,234 / 2,816

By Subject

Subject	Accuracy	Correct / Total
Anaesthesia	41.18%	14 / 34
Anatomy	44.02%	103 / 234
Biochemistry	50.29%	86 / 171
Dental	39.83%	525 / 1,318
ENT	32.08%	17 / 53
Forensic Medicine	37.31%	25 / 67
Gynaecology & Obstetrics	43.30%	97 / 224
Medicine	46.44%	137 / 295
Microbiology	42.62%	52 / 122
Ophthalmology	41.38%	24 / 58
Orthopaedics	30.00%	6 / 20
Pathology	48.96%	165 / 337
Pediatrics	40.60%	95 / 234
Pharmacology	44.03%	107 / 243
Physiology	55.56%	95 / 171
Psychiatry	37.50%	6 / 16
Radiology	53.62%	37 / 69
Skin	29.41%	5 / 17
Social & Preventive Medicine	42.64%	55 / 129
Surgery	41.19%	152 / 369

---

Evaluation NoteBook

---

Technical Specifications

Component	Details
Base Architecture	Qwen/Qwen3-VL-2B-Thinking
Context & Vision	Native multimodal support with Dynamic M-RoPE
Dataset	mint-medmax/medmax_data and synthetic data
Precision	Native float32
Training Platform	Apple Silicon chip
Training Framework	trl with custom fused kernels
Primary Goal	Memory-stable multimodal RL post training and high accuracy benchmarks

---

Benchmark Dataset Citations

1. SLAKE: BoKelvin/SLAKE
2. MedXpertQA: TsinghuaC3I/MedXpertQA
3. VQA-RAD: flaviagiammarino/vqa-rad
4. PubMedQA: qiaojin/PubMedQA
5. MedQA (USMLE): GBaker/MedQA-USMLE-4-options
6. MedMCQA: openlifescienceai/medmcqa
7. Path-VQA: flaviagiammarino/path-vqa

---

Limitations & Safety Guidelines

Research & Educational Use Only

MadriMed 1.2 is an experimental research model. It is not certified for autonomous clinical diagnostic decision-making.

No Treatment Prescriptions

The model should not be used as an authoritative source for:

• Pharmaceutical dosages
• Surgical interventions
• Direct treatment pathways
• Other high-stakes clinical decisions

All model-generated interpretations, medical reasoning, and visual inferences should be reviewed and independently confirmed by a qualified, licensed medical professional before being used in a clinical context.

---

Citation

@software{madrimed1.2-VL-2B,
  title = {MadriMed 1.2: Advanced Multimodal Medical Reasoning at the 2B Scale},
  author = {Madrisight},
  year = {2026},
  url = {https://huggingface.co/madrisight/madrimed1.2-VL-2B}
}