MediX-R1 (reproduction), 4B and 8B

Two Qwen3-VL checkpoints trained with reinforcement learning following the MediX-R1 recipe. They are our reproduction of MediX-R1, the "MediX-R1-4B" and "MediX-R1-8B" rows of the main results table in our paper. Each model is in its own folder.

Folder	Base model	VQA	QA	Macro	Micro
MediX-R1-4B	Qwen/Qwen3-VL-4B-Instruct	49.2	70.5	63.0	53.8
MediX-R1-8B	Qwen/Qwen3-VL-8B-Instruct	49.3	74.4	65.5	56.0

Accuracy (%) over 6 medical VQA and 11 medical text QA benchmarks, greedy decoding, answers scored by three LLM judges. The README in each folder has the training recipe, the prompt format and the per-benchmark results.

Usage

import torch
from huggingface_hub import snapshot_download
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

name = "MediX-R1-8B"   # or "MediX-R1-4B"
path = snapshot_download("Shayan0179/Medix-R1", allow_patterns=f"{name}/*") + f"/{name}"
model = Qwen3VLForConditionalGeneration.from_pretrained(path, dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(path)

PROMPT = open(f"{path}/prompt.txt").read()   # the training instructions, ending with "Question: \n"
messages = [{"role": "user", "content": [
    {"type": "text", "text": PROMPT},
    {"type": "image", "image": "chest_xray.png"},
    {"type": "text", "text": "Is there a pleural effusion?"},
]}]
inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True,
                                       return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

The model answers with a modality tag, then its reasoning in <thinking>...</thinking>, then the final answer in <answer>...</answer>.

License

Apache 2.0, following the base models.