MetaReason-RL

MetaReason-RL is the geometry-reasoning checkpoint from MetaReason. It uses the Qwen3-VL architecture and contains four BF16 safetensors shards.

Load and run

pip install "transformers>=4.57.1" accelerate safetensors pillow

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "pH202411/MetaReason-RL"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": "geometry_problem.png"},
        {"type": "text", "text": "Solve this geometry problem and give the final answer."},
    ],
}]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=4096)

output = output[:, inputs.input_ids.shape[1]:]
print(processor.batch_decode(output, skip_special_tokens=True)[0])